A hybrid exchange intelligent classification method and system based on graph data stream features

Abstract

The application discloses a kind of hybrid exchange intelligent classification method and system based on graph data stream feature, by constructing graph data stream feature portrait model, extract multi-dimensional characteristic parameters such as traffic size, duration, packet interval and graph traversal depth, apply multi-dimensional characteristic threshold determination algorithm to intelligently classify graph data stream, based on classification result through exchange path decision algorithm selects optical circuit switching OCS path or electric packet switching EPS path and carries out hybrid exchange, and performance data is collected to machine learning optimization is carried out to classification threshold and decision rule, finally realizes the dynamic matching and intelligent transmission of graph data stream and network resource.The application solves the technical problems that the existing hybrid exchange system lacks fine identification of graph data stream features, cannot adaptively adjust the exchange strategy, resulting in low network resource utilization efficiency, and realizes efficient intelligent transmission of graph data stream.

Description

Technical Field

[0001] This invention relates to the field of network communication technology, specifically to the field of graph data processing and hybrid switching network technology, and particularly to a hybrid switching intelligent classification method and system based on graph data stream characteristics. Background Technology

[0002] Graph data processing technology has been widely applied in many fields such as big data, artificial intelligence, and network analytics. With the continuous expansion of graph data scale and the increasing complexity of processing requirements, efficient graph data transmission and processing has become a hot research topic. Especially in computing networks, graph data processing faces significant technical challenges.

[0003] Traditional graph data processing network architectures primarily rely on a single switching technology for data transmission. Electrical packet switching (EPS) is widely adopted due to its flexibility and maturity, making it suitable for handling bursty, small-scale data transmissions. Optical circuit switching (OCS), on the other hand, demonstrates significant advantages in large-scale data transmission due to its high bandwidth and low latency. In computing networks, optical switching, with its advantages of high bandwidth, low power consumption, low latency, and high scalability, has become the ideal choice for large-scale traffic transmission in data centers; while electrical switching is suitable for handling small-granular bursty traffic, providing refined flexibility for traditional networks. The evolution of computing networks from electrical switching to hybrid optical-electrical switching has become a consensus.

[0004] In recent years, hybrid switching architectures have begun to be applied in the field of graph data processing. This architecture combines the advantages of EPS and OCS, and theoretically can adapt to the transmission needs of different types of graph data streams. However, existing hybrid switching systems usually use static configuration or simple rules to select the switching method, ignoring the dynamic characteristics and diversity of graph data streams. Especially when faced with the millisecond-level time overhead of OCS configuration, how to efficiently select a suitable switching path has become an urgent problem to be solved.

[0005] The main shortcomings of existing technologies are: on the one hand, they lack a refined identification and classification mechanism for graph data stream characteristics, making it impossible to accurately determine when to use OCS to obtain the best performance; on the other hand, existing systems cannot adaptively adjust the switching strategy according to the actual network environment and data stream changes, resulting in low network resource utilization efficiency, especially when processing graph data with special traversal patterns. Summary of the Invention

[0006] The purpose of this invention is to provide a hybrid switching intelligent classification method and system based on graph data stream features, so as to solve the technical problems of low network resource utilization efficiency caused by the lack of refined identification of graph data stream features and the inability to adaptively adjust switching strategies in existing hybrid switching systems.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A hybrid switching intelligent classification method based on graph data stream features is proposed. The method involves acquiring graph data streams, parsing them, and extracting multi-dimensional feature parameters including traffic volume, duration, packet interval, and graph traversal depth to construct a graph data stream feature profile model. Based on this model, a multi-dimensional feature thresholding algorithm is applied to classify the graph data streams, outputting the classification results. Based on these classification results, a switching path decision algorithm selects either an Optical Circuit Switching (OCS) path or an Electrical Packet Switching (EPS) path as the actual switching path for hybrid switching of the graph data streams. Performance data is collected to evaluate the actual switching effect, and the classification thresholds and decision rules are optimized using machine learning based on this data to update the classification parameters and rules. Finally, according to the updated classification parameters and rules, dynamic matching of the graph data streams with network resources is performed to achieve intelligent transmission of the graph data streams.

[0009] Further, the graph data stream is acquired, parsed, and multi-dimensional feature parameters including traffic size, duration, packet interval, and graph traversal depth are extracted to construct a graph data stream feature profile model. This includes: acquiring the graph data stream; parsing the graph data stream via a network interface to obtain a preliminary graph data stream set; calculating the packet arrival interval distribution, flow duration, and burstiness index based on the preliminary graph data stream set to obtain a traffic time-series feature vector; calculating the number of bytes, number of packets, and average packet size of the traffic time-series feature vector and normalizing them to obtain a scale feature vector; analyzing the graph traversal depth, number of accessed nodes, and edge traversal frequency based on the scale feature vector to obtain a graph structure feature vector; and constructing the graph data stream feature profile model based on the traffic time-series feature vector, the scale feature vector, and the graph structure feature vector using a feature fusion algorithm.

[0010] Furthermore, based on the graph data stream feature profiling model, a multi-dimensional feature threshold determination algorithm is applied to classify the graph data stream and output the graph data stream classification result. This includes: initializing traffic size threshold, duration threshold, and graph traversal depth threshold based on the graph data stream feature profiling model to obtain multi-dimensional determination parameters; configuring feature weights of the multi-dimensional determination parameters according to the network environment and application scenario; calculating a score for the graph data stream features after configuring feature weights to obtain a comprehensive determination score; classifying the graph data stream type according to the comprehensive determination score and preset determination rules to obtain a preliminary classification result; adding confidence information to the preliminary classification result and outputting the graph data stream classification result.

[0011] Furthermore, based on the graph data stream classification results, an optical circuit switching (OCS) path or an electrical packet switching (EPS) path is selected as the actual switching path through a switching path decision algorithm to perform hybrid switching of the graph data streams. This includes: monitoring the resource status of OCS and EPS based on the graph data stream classification results and outputting a network resource status map; and, according to the network resource status map, directing the first graph data stream to the OCS optical path and the second graph data stream to the EPS channel, wherein the data volume of the first graph data stream is greater than the data volume of the second graph data stream.

[0012] Furthermore, performance data for evaluating the actual exchange effect is collected, and the classification threshold and decision rules are optimized by machine learning based on the performance data to update the classification parameters and judgment rules. This includes: collecting performance data after the actual exchange of the graph data stream in real time, determining the evaluation index results; establishing and training a machine learning model, inputting the evaluation index results into the machine learning model, and outputting an optimized parameter set; updating the exchange decision rules based on the optimized parameter set, and generating the updated classification parameters and judgment rules.

[0013] Furthermore, based on the updated classification parameters and judgment rules, dynamic matching of graph data streams with network resources is performed to achieve intelligent transmission of graph data streams. This includes: formulating a graph data stream processing strategy based on the updated classification parameters and judgment rules, wherein resources are reserved for high-priority graph data streams in the graph data stream processing strategy; adjusting the switching path configuration in real time based on the graph data stream processing strategy, performing batch configuration and pre-configuration, and outputting optimized configuration instructions; and performing dynamic matching of graph data streams with network resources based on the optimized configuration instructions to achieve intelligent transmission of graph data streams.

[0014] Furthermore, before performing dynamic matching of graph data streams and network resources, the method further includes: acquiring historical transmission records of the graph data streams, extracting traffic change patterns and resource occupancy characteristics, and fusing them to obtain a graph data stream state feature set; based on the graph data stream state feature set, constructing a Markov state transition model, and applying a reachability analysis algorithm to calculate the state transition probability to obtain traffic state prediction results at each time point; based on the traffic state prediction results and the existing network resource status, calculating the resource demand within a future time window; based on the resource demand within the future time window, performing path planning in conjunction with resource constraints of OCS and EPS to obtain a preset resource allocation strategy; and based on the preset resource allocation strategy, collaboratively optimizing the resource demands of multiple data streams to achieve dynamic matching of the graph data streams and network resources.

[0015] Furthermore, the method also includes: acquiring the transmission behavior of the graph data stream, collecting traffic changes, resource usage, and performance index data; based on the traffic changes, resource usage, and performance index data, applying time series analysis methods to extract behavioral patterns, and identifying the inter-flow influence relationships in the behavioral patterns through Granger causality tests to obtain a preliminary causal association network; using the preliminary causal association network to calculate the influence propagation delay and propagation intensity, and outputting a weighted time-series causal graph; identifying key influence paths and propagation chains in the weighted time-series causal graph to obtain a causal propagation model; and predicting the chain reaction of decision adjustment based on the causal propagation model to achieve optimal decision-making for multi-flow collaboration.

[0016] Furthermore, the method also includes: acquiring historical decision records of the graph data stream, extracting decision scenarios and effect data, constructing a decision tree model to characterize decision rules based on the decision scenarios and effect data, and obtaining a historical decision model; generating multiple sets of hypothetical scenarios based on the historical decision model using the Monte Carlo method to generate a counterfactual scenario set; performing simulation experiments to evaluate the effects of different decisions based on the counterfactual scenario set, and outputting a comparison result of decision effects; calculating the optimal possible performance of each decision point in the comparison result of decision effects, and updating the decision rules and evaluation criteria based on the optimal possible performance to achieve continuous optimization of the decision mechanism.

[0017] This invention also provides a hybrid switching intelligent classification system based on graph data stream features, comprising: an acquisition module for acquiring graph data streams, parsing the graph data streams, extracting multi-dimensional feature parameters including traffic size, duration, packet interval, and graph traversal depth, and constructing a graph data stream feature profile model; a classification module for classifying the graph data streams based on the graph data stream feature profile model using a multi-dimensional feature threshold determination algorithm, and outputting the graph data stream classification result; a hybrid switching module for selecting an optical circuit switching (OCS) path or an electrical packet switching (EPS) path as the actual switching path for hybrid switching of the graph data streams based on the graph data stream classification result using a switching path decision algorithm; an optimization module for collecting performance data to evaluate the actual switching effect, and performing machine learning optimization on the classification threshold and decision rules based on the performance data to update the classification parameters and determination rules; and a transmission module for performing dynamic matching of the graph data streams with network resources according to the updated classification parameters and determination rules to achieve intelligent transmission of the graph data streams.

[0018] The beneficial effects of this invention are as follows:

[0019] By constructing a multi-dimensional graph data stream feature profiling model, key parameters such as traffic size, duration, packet interval, and graph traversal depth are extracted, enabling accurate characterization of graph data stream features and providing a reliable basis for subsequent intelligent classification.

[0020] An adaptive multidimensional feature threshold determination mechanism was designed, which can intelligently distinguish between different types of data streams such as "large streams" and "small streams", "continuous graph traversal operations" and "short-term control signaling", thereby improving the accuracy and adaptability of classification.

[0021] A dynamic mapping relationship between data flow characteristics and switching methods was established. Based on traffic characteristics, the OCS or EPS path was intelligently selected, making full use of the advantages of different switching technologies and improving the efficiency of network resource utilization.

[0022] By introducing a machine learning feedback optimization mechanism, the classification threshold and decision rules are dynamically adjusted through continuous learning of actual exchange effects, thereby achieving continuous improvement in classification accuracy and continuous optimization of system performance.

[0023] The innovation realizes an efficient orchestration method for flow-path collaboration. Through dynamic resource reservation, batch configuration and other technical means, it effectively reduces the performance impact caused by the millisecond-level configuration overhead of OCS and realizes efficient and intelligent transmission of graph data streams. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0025] Figure 2 A schematic diagram illustrating the process of building a feature profiling model for graph data streams;

[0026] Figure 3 This is a schematic diagram of the multidimensional feature threshold determination process;

[0027] Figure 4 A schematic diagram of the intelligent selection and configuration process for switching paths;

[0028] Figure 5 A flowchart illustrating the process of optimizing feedback mechanisms for machine learning;

[0029] Figure 6 This is a schematic diagram of the efficient orchestration architecture for flow-path collaboration. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1

[0032] like Figure 1As shown, this embodiment provides a hybrid exchange intelligent classification method based on graph data stream features, including:

[0033] S1: Obtain the graph data stream, parse the graph data stream, extract multi-dimensional feature parameters including traffic size, duration, data packet interval and graph traversal depth, and construct a graph data stream feature profile model.

[0034] In this embodiment, raw graph data streams are captured via a network interface. These graph data streams originate from application scenarios such as graph database query operations, graph algorithm computation tasks, or graph neural network training processes. The captured raw data streams contain a large number of network data packets, each carrying information such as source address, destination address, port number, timestamp, and payload.

[0035] The captured data packets are subjected to deep parsing to extract data flow identification information, including the five-tuple information (source IP address, destination IP address, source port, destination port, and transport protocol). Based on this identification information, data packets belonging to the same session or connection are grouped and aggregated to form a preliminary set of graph data flows. For example, all data packets from the same graph database query task are identified as a single graph data flow.

[0036] For each graph data stream, a timing characteristic analysis is performed. The arrival time interval between adjacent data packets is calculated, and the distribution characteristics of these intervals, such as average interval, variance, maximum and minimum intervals, are statistically analyzed to characterize the burstiness of the data stream. Simultaneously, the time span from the first data packet to the last data packet is recorded as a stream duration parameter. These timing characteristics can reflect the transmission pattern of the graph data stream. For example, continuous large-scale graph traversal operations typically exhibit longer durations and relatively stable packet intervals, while short-term control signaling exhibits shorter durations and larger interval fluctuations.

[0037] The scale characteristics of the computational graph data stream are determined. This involves counting the total number of bytes in the entire data stream, reflecting the magnitude of data transmission; counting the total number of packets, reflecting the frequency of transmission; and calculating the average size of each packet. To eliminate the impact of bandwidth differences under different network environments, these scale characteristics are normalized, for example, by dividing the number of bytes by the network link bandwidth and standardizing the duration to a uniform time scale, resulting in a scale feature vector.

[0038] Analyze graph data stream content to identify unique operational patterns in graph data processing. By parsing application layer protocols or using machine learning-based traffic classification techniques, identify the depth of graph traversal operations, i.e., the maximum number of hops from the starting node; count the number of graph nodes visited to reflect the coverage of queries or computations; and analyze the frequency distribution of edge traversals to identify hotspot paths and access patterns. These graph structural features can distinguish different types of graph operations; for example, depth-first search exhibits a larger traversal depth and relatively fewer node visits, while breadth-first search exhibits a shallower traversal depth but visits a large number of nodes.

[0039] The extracted time-series feature vectors, scale feature vectors, and graph structure feature vectors are fused. Feature fusion algorithms, such as weighted concatenation, principal component analysis dimensionality reduction, or deep neural network autoencoding, are employed to integrate multiple feature vectors into a unified high-dimensional feature representation, constructing a graph data stream feature profiling model. This model comprehensively characterizes the graph data stream's properties in terms of time, scale, and structure, providing rich feature data for subsequent intelligent classification.

[0040] S2: Based on the graph data stream feature profiling model, a multi-dimensional feature threshold determination algorithm is applied to classify the graph data stream and output the graph data stream classification result.

[0041] In this embodiment, the classification system initializes a set of key feature threshold parameters based on the constructed graph data stream feature profiling model. These threshold parameters include a traffic size threshold T1, used to distinguish between large and small traffic streams; for example, T1 is set to 100MB, and data streams exceeding this threshold are initially identified as large streams. A duration threshold T2 is used to distinguish between long-term continuous transmission and short-term burst transmission; for example, T2 is set to 10 seconds. A graph traversal depth threshold T3 is used to distinguish between deep graph operations and shallow graph operations; for example, T3 is set to 5 hops. These initial thresholds can be set based on historical statistical data and domain expert experience.

[0042] Based on the actual network environment characteristics and application scenario requirements, importance weights are assigned to each dimension of features. For example, in network environments with limited bandwidth resources, the weight of the traffic volume feature can be set higher to prioritize high-volume transmission; in latency-sensitive application scenarios, the weight of the duration feature can be increased to quickly identify data streams requiring low-latency transmission; in graph computation-intensive scenarios, the weight of the graph traversal depth feature is more critical. Weight configuration can be determined using the analytic hierarchy process (AHP), entropy weighting, or optimization algorithms based on historical performance data.

[0043] A score is calculated for the feature profile of each graph data stream. The feature values ​​of each dimension are compared with their corresponding thresholds to obtain a score for each dimension. For example, a flow size exceeding T1 receives 1 point, otherwise 0 points; a duration exceeding T2 receives 1 point, otherwise 0 points; and a graph traversal depth exceeding T3 receives 1 point, otherwise 0 points. Then, each dimension's score is multiplied by its corresponding weight, and the results are summed to obtain a comprehensive score. For example, if the weight for flow size is 0.5, the weight for duration is 0.3, and the weight for graph traversal depth is 0.2, and a data stream scores 1, 1, and 0 in the three dimensions respectively, then the comprehensive score is 0.5 x 1 + 0.3 x 1 + 0.2 x 0 = 0.8.

[0044] Based on preset judgment rules, graph data streams are categorized. For example, data streams with a comprehensive judgment score greater than 0.7 are classified as "large flow - continuous graph traversal operation," suitable for transmission using OCS; those with a comprehensive judgment score between 0.3 and 0.7 are classified as "medium flow," requiring further refinement; and those with a comprehensive judgment score less than 0.3 are classified as "small flow - short-term control signaling," suitable for transmission using EPS. This multi-dimensional comprehensive judgment allows for a more accurate identification of the transmission requirements of data streams.

[0045] Add confidence information to the classification results. Confidence reflects the reliability of the classification decision and can be calculated based on factors such as the distance between feature values ​​and thresholds, and historical classification accuracy statistics. For example, if a data stream has a comprehensive decision score of 0.85, which is much higher than the classification boundary of 0.7, its confidence in being classified as a "major flow" is high and can be set to 0.95; if the comprehensive decision score is 0.72, which is close to the classification boundary, the confidence is low and can be set to 0.6. The output graph data stream classification results contain three parts: data stream identifier, classification type, and confidence score, providing a basis for subsequent exchange path decisions.

[0046] S3: Based on the graph data stream classification results, the optical circuit switching (OCS) path or electrical packet switching (EPS) path is selected as the actual switching path through the switching path decision algorithm to perform the mixed switching of the graph data stream.

[0047] In this embodiment, the switching decision system intelligently selects the optimal switching path based on the graph data stream classification results and the real-time status of network resources. First, the monitoring system monitors the resource status of OCS and EPS in real time, including the number of available wavelengths of OCS optical paths, the bandwidth occupancy of established optical paths, the latency overhead of optical path establishment and teardown, and performance indicators such as link bandwidth utilization, queue length, packet loss rate, and end-to-end latency of EPS channels, and constructs a network resource status graph.

[0048] Based on the data flow type and confidence information in the classification results, switching path decision rules are applied. For data flows classified as "large-flow - persistent graph traversal operation" with high confidence, the OCS optical path is preferentially selected for transmission. The dedicated optical path provided by OCS has the characteristics of high bandwidth, low latency, and no congestion, which can meet the performance requirements of large-scale graph data transmission. When selecting the OCS path, the optimal optical path configuration scheme is calculated, including selecting a suitable wavelength, determining the optical path routing, and estimating the optical path establishment time. Then, control commands are sent to the optical switching equipment to establish an end-to-end optical path connection.

[0049] For data flows classified as "small flow - short-term control signaling," the EPS channel is selected for transmission. EPS employs a packet forwarding mechanism, offering advantages such as flexible configuration and rapid response, making it suitable for handling small-scale, bursty data transmission needs. When selecting the EPS path, the packet forwarding strategy is optimized, calculating the shortest or optimal path based on network topology and link status, and setting corresponding quality of service parameters, such as priority, bandwidth reservation, and queue scheduling strategies, to ensure timely and reliable data flow transmission.

[0050] For data streams with low classification confidence or those on the classification boundary, a hybrid decision-making strategy is adopted. Initial transmission can be performed via EPS while monitoring the actual transmission characteristics of the data stream. If the actual size and duration of the data stream exceed expectations, the system dynamically switches to the OCS path. Alternatively, a parallel transmission strategy can be used, employing both OCS and EPS for transmission simultaneously, dynamically adjusting the traffic allocation ratio between the two paths based on real-time performance feedback.

[0051] The switching decision system translates the decision results into specific network configuration instructions and sends them to the corresponding switching devices. For OCS, the configuration instructions include port connection relationships of the optical switching matrix, wavelength allocation information, etc.; for EPS, the configuration instructions include routing table entries, forwarding rules, QoS policies, etc. The execution of the configuration instructions completes the actual establishment of the switching path, and the graph data stream begins to be transmitted through the selected path.

[0052] S4: Collect performance data to evaluate the actual exchange effect, and perform machine learning optimization on the classification threshold and decision rules based on the performance data to update the classification parameters and decision rules.

[0053] In this embodiment, performance data of the graph data stream is collected in real time during the actual switching and transmission process. This performance data includes throughput, i.e., the amount of data successfully transmitted per unit time, reflecting transmission efficiency; end-to-end latency, i.e., the time it takes for data to be sent from the source end to the destination end, reflecting transmission timeliness; packet loss rate, i.e., the proportion of data packets lost during transmission to the total number of data packets, reflecting transmission reliability; resource utilization, including the bandwidth utilization of the OCS optical path and the bandwidth utilization of the EPS link, reflecting resource utilization efficiency; and configuration time overhead, especially the time required for OCS optical path establishment and teardown, reflecting switching flexibility.

[0054] The collected performance data is correlated with the classification results to evaluate classification accuracy. The actual performance of data streams classified as "large flows" and selected for OCS transmission is compared. If these data streams do indeed exhibit high throughput and long duration, and OCS transmission brings a significant performance improvement, then the classification is accurate. Conversely, if some data streams classified as "large flows" are actually small in scale or short in duration, and using OCS causes performance degradation due to configuration overhead, then there is a misclassification. Statistics such as classification accuracy, false positive rate, and false negative rate are calculated to identify classification scenarios and parameters that need optimization.

[0055] A machine learning model is built to learn the complex relationship between data stream features and optimal exchange methods. The construction process of the machine learning model is as follows: First, a training dataset is prepared, including feature profile data of historical graph data streams, the actual selected exchange paths, and the corresponding performance evaluation results. The performance evaluation results are quantified into labels, for example, high throughput, low latency, and high resource utilization are marked as "excellent," and vice versa as "poor."

[0056] Choose a suitable machine learning algorithm, such as decision trees, random forests, gradient boosting trees, or deep neural networks. Taking random forests as an example, this algorithm can handle multi-dimensional feature inputs, automatically learn non-linear relationships between features, and has good generalization ability. Construct a random forest model and set model parameters, such as setting the number of decision trees to 100, the maximum depth of each tree to 10 layers, and the feature sampling ratio to 0.8.

[0057] The model is trained using a training dataset. Data stream feature profiles are used as input features, and performance evaluation labels are used as output targets. The model learns the mapping relationship between features and performance during training. During training, the model automatically adjusts its internal parameters and optimizes the classification boundary, enabling it to accurately predict the optimal swapping method based on the feature profiles. After training, the model's performance is evaluated using a validation dataset, calculating metrics such as prediction accuracy, recall, and F1 score to ensure the model has good predictive capabilities.

[0058] The trained machine learning model is applied to the actual system, taking the current performance evaluation data and classification results as input, and the model outputs an optimized parameter set. The optimized parameter set includes adjusted feature thresholds, such as adjusting the traffic size threshold T1 from 100MB to 120MB and the duration threshold T2 from 10 seconds to 8 seconds; and adjusted feature weights, such as increasing the weight of the graph traversal depth feature and decreasing the weight of the packet interval feature.

[0059] Based on the optimized parameter set, the classification system's decision rules are updated. New threshold and weight parameters are applied to the multi-dimensional feature thresholding algorithm, while the swapping decision rules are updated, such as adjusting the swapping path selection strategy for different classification types and optimizing the confidence calculation method. The updated classification parameters and decision rules can more accurately identify the characteristics of graph data streams and make better swapping path decisions, thereby continuously improving the overall system performance.

[0060] S5: Based on the updated classification parameters and judgment rules, perform dynamic matching of graph data streams and network resources to achieve intelligent transmission of graph data streams.

[0061] In this embodiment, the flow-path collaborative orchestration system formulates differentiated graph data flow processing strategies based on updated classification parameters and judgment rules. For different types of data flows, corresponding caching strategies, priority strategies, and resource reservation strategies are designed. For example, for high-priority "large flow - persistent graph traversal operations," OCS optical path resources are pre-allocated before the data flow arrives to avoid configuration delays caused by temporary optical path establishment; for low-priority "small flow - short-term control signaling," a best-effort transmission strategy is adopted, utilizing the idle bandwidth of the EPS for transmission.

[0062] Real-time monitoring of network status and data flow changes allows for dynamic adjustment of switching path configurations. When network congestion, link failures, or resource shortages are detected, the selection of data flow switching paths is promptly reassessed, and path switching or traffic migration is performed as necessary. For example, if an OCS optical path is interrupted due to a fault, the affected data flow is quickly switched to a backup optical path or temporarily used for transmission via an EPS channel to ensure service continuity.

[0063] To address the millisecond-level configuration overhead of OCS, a batch configuration and pre-configuration mechanism is designed. Multiple data streams requiring OCS optical path establishment are aggregated and configured uniformly, reducing configuration frequency and overall overhead. Simultaneously, based on a data stream prediction model, potentially large-volume data streams are predicted in advance, and optical path connections are pre-established, allowing data streams to immediately utilize existing optical paths upon arrival, eliminating configuration waiting time.

[0064] Optimized configuration commands are generated, including commands for establishing, adjusting, and tearing down OCS optical paths, commands for updating EPS routing tables, and commands for configuring QoS parameters. These commands are optimized and orchestrated to ensure efficient execution of configuration operations and rational allocation of resources. For example, some OCS optical paths are pre-established during off-peak hours to quickly respond to data flow demands during peak hours; and optical paths are promptly torn down when idle, freeing up wavelength resources for other data flows.

[0065] The system dynamically matches graph data streams with network resources, connecting data stream transmission needs with available network resources in real time. Through intelligent scheduling algorithms, data streams are allocated to the most suitable transmission paths, maximizing resource utilization and optimizing transmission performance. The entire process forms a closed-loop control, continuously iterating and improving from data stream feature extraction, intelligent classification, path decision-making, performance monitoring to parameter optimization, achieving efficient and intelligent transmission of graph data streams.

[0066] Example 2

[0067] This embodiment is a further refinement of S1 in Embodiment 1, such as... Figure 2 As shown, the process involves acquiring a graph data stream, parsing the graph data stream, extracting multi-dimensional feature parameters including traffic size, duration, packet interval, and graph traversal depth, and constructing a graph data stream feature profile model. This includes: acquiring the graph data stream; parsing the graph data stream via a network interface to obtain a preliminary graph data stream set; calculating the packet arrival interval distribution, flow duration, and burstiness index based on the preliminary graph data stream set to obtain a traffic temporal feature vector; calculating the number of bytes, number of packets, and average packet size of the traffic temporal feature vector and normalizing them to obtain a scale feature vector; analyzing the graph traversal depth, number of accessed nodes, and edge traversal frequency based on the scale feature vector to obtain a graph structure feature vector; and constructing the graph data stream feature profile model based on the traffic temporal feature vector, the scale feature vector, and the graph structure feature vector using a feature fusion algorithm.

[0068] S1.1: Obtain the graph data stream, and parse the graph data stream through the network interface to obtain a preliminary graph data stream set.

[0069] In this embodiment, the data capture module captures all network data packets passing through a key network node in real time using monitoring devices deployed at that node. The monitoring device can be a dedicated network traffic collector, a mirror port of a switch, or a monitoring interface of a software-defined network controller. The captured raw data packets are stored in PCAP format, containing complete packet header and payload information.

[0070] The captured data packets are parsed to extract key fields from the network and transport layers. The source and destination IP addresses are extracted from the IP header, and the source and destination ports are extracted from the TCP or UDP header to identify the transport protocol type. This information forms a five-tuple identifier for the data stream, uniquely identifying it. For example, all data packets with a source IP of 192.168.1.100, a destination IP of 10.0.0.50, a source port of 12345, a destination port of 8080, and a TCP protocol are identified as belonging to the same data stream.

[0071] Based on the 5-tuple identifier, data packets are grouped and aggregated. A data flow session table is established to record information such as the first packet time, last packet time, and data packet sequence for each data flow. For TCP connections, the start and end of the data flow can be accurately identified based on the three-way handshake and four-way handshake process; for UDP transmission, a timeout threshold is set based on the time interval between data packet arrivals, for example, if no new data packets arrive within 30 seconds, the data flow is considered to have ended.

[0072] This creates a preliminary set of graph data streams, each containing its identifier, detailed records of all data packets, and basic statistical information such as the total number of data packets, total number of bytes, and duration. This preliminary set provides the foundational data for subsequent feature extraction and analysis.

[0073] S1.2: Based on the preliminary graph data stream set, calculate the data packet arrival interval distribution, stream duration, and burst index to obtain the traffic time-series feature vector.

[0074] In this embodiment, the time-series feature extraction module performs detailed time-dimensional analysis on each graph data stream. First, it calculates the arrival time interval between adjacent data packets. It then iterates through all data packets in the data stream, records the arrival timestamp of each data packet, and calculates the time difference between the i-th data packet and the (i+1)-th data packet to obtain a series of interval values.

[0075] Analyze the distribution characteristics of these interval values. Calculate the average interval value, reflecting the average frequency of data packet arrival; calculate the standard deviation or variance of the interval, reflecting the degree of fluctuation in the arrival frequency; calculate the maximum and minimum values ​​of the interval to identify burst transmissions and idle periods; plot the interval distribution histogram or probability density function to analyze the morphological characteristics of the interval distribution, such as whether it conforms to a Poisson distribution, a normal distribution, or other specific distribution patterns.

[0076] Calculate the stream duration, which is the time span from the arrival of the first data packet to the arrival of the last data packet. Stream duration reflects the lifespan of the data stream. Continuous graph traversal operations typically have a long duration, which may reach several seconds or even minutes; while short-term control signaling has a very short duration, usually in the millisecond range or hundreds of milliseconds.

[0077] Burst indexes are used to quantify the burst transmission characteristics of data streams. A burst index can be defined as the ratio of the peak to the average number of data packets or bytes per unit of time. For example, dividing the data stream into multiple time windows, each 100 milliseconds long, counting the number of data packets within each window, and calculating the ratio of the maximum number of packets in the maximum window to the average number of packets in the average window yields the burst index. A high burst index indicates significant burst transmission behavior in the data stream, while a low burst index indicates relatively stable data stream transmission.

[0078] The calculated statistical characteristics of packet arrival intervals, flow duration, burstiness indicators, and other parameters are combined to form a traffic time-series feature vector. This vector comprehensively characterizes the transmission behavior of graph data streams in the time dimension.

[0079] S1.3: Calculate the number of bytes, number of data packets, and average packet size of the traffic time-series feature vector and perform normalization processing to obtain the scale feature vector.

[0080] In this embodiment, the scale feature extraction module performs quantitative analysis on the transmission scale of the graph data stream. The total number of bytes in the statistical data stream, i.e., the sum of the payload bytes of all data packets, reflects the total amount of data transmitted. For example, a large-scale graph traversal operation may transmit several gigabytes of data, while a simple query request may only transmit a few kilobytes.

[0081] The total number of data packets in a data stream is the number of packets contained within the stream. The combination of the number of data packets and the total number of bytes reflects the granularity of data transmission. For data streams with the same number of bytes, a higher number of data packets indicates that each packet is small, potentially representing control signaling or small message transmissions; conversely, a lower number of data packets indicates that each packet is large, potentially representing large data transmissions.

[0082] Calculate the average packet size, which is the total number of bytes divided by the number of data packets. Average packet size is an important characteristic for distinguishing data stream types. Graph data transmission typically uses larger data packets to improve transmission efficiency, and the average packet size may be close to the MTU value, such as 1500 bytes; while control protocol messages are usually smaller, and the average packet size may only be tens to hundreds of bytes.

[0083] To eliminate dimensional differences across different network environments and application scenarios, scale characteristics are normalized. Normalization methods can include max-min normalization, mapping characteristic values ​​to the range of 0 to 1; or standardization, converting characteristic values ​​into a distribution with a mean of 0 and a standard deviation of 1. For example, dividing the total number of bytes by the maximum transmission capacity of the network link and dividing the duration by a preset maximum observation time window yields the normalized scale characteristic values.

[0084] The normalized parameters, such as total bytes, number of data packets, and average packet size, are combined to form a scale feature vector. This vector represents the transmission scale characteristics of the graph data stream in a standardized form, facilitating subsequent feature fusion and classification.

[0085] S1.4: Based on the scale feature vector, analyze the graph traversal depth, the number of visited nodes, and the edge traversal frequency to obtain the graph structure feature vector.

[0086] In this embodiment, the graph structure feature recognition module performs in-depth analysis of the graph data stream content to extract operation mode features unique to graph data processing. This process requires parsing the application layer protocol or using deep packet inspection technology to identify the application type and operation semantics of the data stream.

[0087] Identifying graph traversal depth. For graph database query operations, traversal depth information is extracted by parsing the query statement or API call parameters. For example, path length limits in the Cypher query language and the number of steps in the Gremlin query directly reflect the graph traversal depth. For graph algorithm computation tasks, traversal depth is inferred by analyzing the algorithm type and parameters, such as the K value in a K-hop neighbor query and the number of iterations in the PageRank algorithm. Graph traversal depth is a key feature distinguishing deep graph operations from shallow graph operations. Deep traversal typically requires transmitting more intermediate result data, placing higher demands on network bandwidth and latency.

[0088] Count the number of accessed nodes. Analyze the graph data content transmitted in the data stream, identify the involved graph node identifiers, and count the number of different nodes. The number of accessed nodes reflects the coverage of graph operations. Large-scale graph analysis tasks may access millions or even billions of nodes, while small-scale local queries may only access a few or dozens of nodes. The number of nodes is usually positively correlated with the amount of data transmitted and is an important basis for judging the scale of the data stream.

[0089] Analyze edge traversal frequency. By statistically analyzing the number of times graph edges are visited in the data stream, frequently visited hotspot edges and access patterns can be identified. The distribution of edge traversal frequency reflects the access characteristics of graph operations; for example, edges of certain central nodes may be frequently visited, forming a star-shaped access pattern; or specific paths may be repeatedly traversed, forming a chain-like access pattern. Edge traversal frequency information helps predict the subsequent behavior and resource requirements of the data stream.

[0090] The extracted graph traversal depth, number of visited nodes, edge traversal frequency, and other parameters are quantized and normalized to form a graph structure feature vector. For cases where graph structure information cannot be directly extracted from the data stream content, machine learning-based traffic classification techniques can be used to train a classifier to identify the application type of the graph data stream and indirectly infer its graph structure features.

[0091] S1.5: Based on the traffic time-series feature vector, the scale feature vector, and the graph structure feature vector, the graph data flow feature profile model is constructed through a feature fusion algorithm.

[0092] In this embodiment, the feature fusion module integrates the multiple feature vectors extracted above to construct a unified graph dataflow feature profiling model. The goal of feature fusion is to effectively combine feature information of different dimensions and scales to form a comprehensive, compact, and discriminative feature representation.

[0093] A weighted concatenation method is used for feature fusion. The traffic time-series feature vector, scale feature vector, and graph structure feature vector are concatenated sequentially to form a high-dimensional feature vector. For example, the time-series feature vector contains 5 elements, the scale feature vector contains 3 elements, and the graph structure feature vector contains 4 elements; after concatenation, a 12-dimensional comprehensive feature vector is obtained. To balance the contributions of different feature types, weight coefficients can be assigned to each feature component; for example, the time-series feature has a weight of 0.3, the scale feature has a weight of 0.4, and the graph structure feature has a weight of 0.3. Each feature vector is multiplied by its corresponding weight before concatenation.

[0094] To reduce feature dimensionality, eliminate feature redundancy, and extract key information, principal component analysis (PCA) can be used for dimensionality reduction. PCA is performed on the concatenated high-dimensional feature vectors to calculate the covariance matrix of the features, solve for the eigenvalues ​​and eigenvectors, and select the first few principal components as the feature representation after dimensionality reduction. For example, 12-dimensional features can be reduced to 5-dimensionality while retaining more than 95% of the information, reducing computational complexity while maintaining the discriminative power of the features.

[0095] Feature fusion can also be achieved using deep neural network autoencoders. An autoencoder network is constructed, consisting of an encoder and a decoder. The encoder compresses the input multidimensional feature vector layer by layer, learning a compact feature representation; the decoder restores the compressed feature representation to the original feature vector. By training the autoencoder to minimize the reconstruction error, the intermediate layer features learned by the encoder become the fused feature representation. This method can automatically learn the nonlinear relationships between features, extracting deeper semantic information.

[0096] The constructed graph data stream feature profiling model is represented in vector form, with each dimension corresponding to a feature parameter. The values ​​of the vectors comprehensively characterize the graph data stream's features in terms of time, scale, and structure. This model serves as input for subsequent classification decisions, providing rich and accurate feature data for intelligent classification.

[0097] Example 3

[0098] This embodiment is a further refinement of S2 in Embodiment 1, such as... Figure 3As shown, based on the graph data stream feature profiling model, a multi-dimensional feature threshold determination algorithm is applied to classify the graph data stream and output the graph data stream classification result. This includes: initializing traffic size threshold, duration threshold, and graph traversal depth threshold based on the graph data stream feature profiling model to obtain multi-dimensional determination parameters; configuring feature weights of the multi-dimensional determination parameters according to the network environment and application scenario; calculating a score for the graph data stream features after configuring feature weights to obtain a comprehensive determination score; classifying the graph data stream type according to the comprehensive determination score and preset determination rules to obtain a preliminary classification result; adding confidence information to the preliminary classification result and outputting the graph data stream classification result.

[0099] S2.1: Based on the graph data stream feature profiling model, initialize the traffic size threshold, duration threshold, and graph traversal depth threshold to obtain multi-dimensional judgment parameters.

[0100] In this embodiment, the threshold initialization module sets initial threshold parameters for key features based on historical data statistics and expert experience. First, it analyzes the transmission records of historical graph data streams and statistically analyzes the characteristic distribution of different types of data streams. For example, it collects the traffic size data of all graph data streams in the past month and draws a histogram of traffic size distribution. It finds that most small traffic data streams are concentrated below 10MB, while large traffic data streams are mainly distributed above 100MB, with a clear dividing region in between.

[0101] Based on the statistical analysis results, a threshold T1 is set for the data flow size. Cluster analysis can be used to divide the data flow into two or more categories according to the flow size, with the midpoint between the cluster centers serving as the classification threshold. For example, if K-means clustering reveals that the cluster centers for small flow are at 5MB and the cluster centers for large flow are at 200MB, then T1 can be set to 100MB as the initial threshold for distinguishing between small and large flow.

[0102] Similarly, by analyzing the duration distribution of historical data streams, a duration threshold T2 is set. Statistical analysis reveals that the duration of short-time control signaling is typically less than 1 second, while the duration of persistent graph traversal operations typically exceeds 10 seconds. Therefore, T2 can be initially set to 5 seconds as a threshold to distinguish between short-time and persistent transmissions.

[0103] Analyzing the distribution characteristics of graph traversal depth, a threshold T3 is set for graph traversal depth. Shallow graph operations, such as one-hop neighbor queries and two-hop relationship queries, typically have a traversal depth of 3 or less; deep graph operations, such as multi-hop path searches and full graph traversals, may reach a traversal depth of 5 hops or more. Based on application scenario requirements and performance analysis, T3 is initially set to 4 hops as the threshold for distinguishing between shallow and deep graph operations.

[0104] In addition to the three main thresholds mentioned above, auxiliary thresholds can be set for other feature parameters, such as the number of data packets, the average packet size, and the burstiness threshold, forming a complete set of multi-dimensional decision parameters. These initial threshold parameters provide a benchmark for classification and can be dynamically adjusted through machine learning optimization.

[0105] S2.2: Configure the feature weights of the multidimensional judgment parameters according to the network environment and application scenario.

[0106] In this embodiment, the weight configuration module assigns importance weights to features of each dimension based on the actual network environment characteristics and application scenario requirements. Weight configuration needs to comprehensively consider multiple factors, including network resource status, service priority, and performance goals.

[0107] Analyze the current network resource status. If network bandwidth is limited and the number of OCS optical paths is also limited, more stringent screening of data flows suitable for OCS usage is required. In this case, the weight of the traffic size characteristic should be set higher, such as 0.5 or 0.6, to ensure that only truly high-volume data flows use valuable OCS resources. If network bandwidth is sufficient and OCS resources are relatively abundant, the weight of the traffic size characteristic can be appropriately reduced, while the weight of other characteristics such as duration and graph traversal depth can be increased to more comprehensively consider data flow characteristics.

[0108] Consider the performance requirements of the application scenario. For latency-sensitive real-time graph analytics applications, such as real-time recommendations on social networks and real-time queries for financial risk control, the weights of duration and burstiness features should be increased to quickly identify data streams requiring low-latency transmission and prioritize the allocation of OCS resources to them. For throughput-sensitive batch graph computation tasks, such as offline graph mining and large-scale graph training, the weights of traffic size and graph traversal depth features are more important.

[0109] The Analytic Hierarchy Process (AHP) is used for weight allocation. A hierarchical structure of feature importance is constructed, with the top layer representing the classification objective, the middle layer representing evaluation criteria such as resource utilization, transmission performance, and configuration overhead, and the bottom layer representing the various feature dimensions. The relative importance of each criterion and the contribution of each feature to each criterion are determined through expert scoring or pairwise comparisons, and finally, the comprehensive weight of each feature is calculated.

[0110] Weight configuration can also be achieved using optimization algorithms based on historical performance data. This involves collecting historical classification decision data and actual transmission performance data, establishing a relationship model between weight parameters and system performance indicators, and then using optimization algorithms such as genetic algorithms and particle swarm optimization to search for the optimal combination of weight parameters that maximizes system performance.

[0111] After configuration, the weight coefficients for each feature are obtained, such as traffic size weight 0.4, duration weight 0.3, graph traversal depth weight 0.2, and other feature weights 0.1. These weight coefficients are used in subsequent scoring calculations and affect the final classification result.

[0112] S2.3: Calculate a score for the graph data stream features after configuring feature weights to obtain a comprehensive judgment score.

[0113] In this embodiment, the feature profiles of each graph data stream are quantitatively scored, and a unified judgment score is obtained by integrating multi-dimensional feature information. The scoring calculation process includes two steps: single-dimensional scoring and multi-dimensional weighted summarization.

[0114] Single-dimensional scoring is performed. For each feature dimension, the feature value is compared with the corresponding threshold, and a score is assigned based on the comparison result. The simplest scoring method is binary scoring, where a feature value exceeding the threshold receives 1 point, and otherwise receives 0 points. For example, if the size of a data stream is 150MB, exceeding the threshold T1 (100MB), then the size dimension scores 1; if the duration is 3 seconds, not exceeding the threshold T2 (5 seconds), then the duration dimension scores 0.

[0115] To more precisely reflect the relative relationship between feature values ​​and thresholds, a continuous scoring method can be used. A scoring function is defined, calculating the score based on the distance between the feature value and the threshold. For example, using the Sigmoid function, the score is close to 1 when the feature value is much larger than the threshold, close to 0 when the feature value is much smaller than the threshold, and around 0.5 when the feature value is close to the threshold. Specifically, the flow rate score can be calculated as 1 divided by 1 plus the negative of the natural constant (flow rate minus T1), divided by the exponent of the scaling parameter. The steepness of the scoring function is controlled by adjusting the scaling parameter.

[0116] For certain features, it may be necessary to consider bidirectional thresholding or interval determination. For example, the ideal range for packet intervals may be within a certain interval; too small or too large is not ideal. In this case, an interval scoring function can be defined, where feature values ​​within the ideal interval receive a high score, and those deviating from the interval receive a low score.

[0117] Perform a multi-dimensional weighted summary. Multiply the individual scores of each dimension by their corresponding weight coefficients, and then sum them to obtain the comprehensive judgment score. The calculation formula is: the comprehensive judgment score equals the traffic size score multiplied by the traffic size weight, plus the duration score multiplied by the duration weight, plus the graph traversal depth score multiplied by the graph traversal depth weight, plus the sum of the scores of all other dimensions multiplied by their corresponding weights.

[0118] For example, if the scores for each dimension of a data stream are: flow size score 0.9, duration score 0.6, graph traversal depth score 0.8, and other feature score 0.5, with corresponding weights of 0.4, 0.3, 0.2, and 0.1, then the comprehensive judgment score is equal to 0.9*0.4+0.6*0.3+0.8*0.2+0.5*0.1=0.75.

[0119] The overall score is a value between 0 and 1. A higher value indicates that the data stream is more in line with the characteristics of "large flow - continuous graph traversal operation" and is more suitable for transmission using OCS; a lower value indicates that the data stream is more in line with the characteristics of "small flow - short-time control signaling" and is more suitable for transmission using EPS.

[0120] S2.4: Based on the comprehensive judgment score, the graph data stream type is divided according to the preset judgment rules to obtain the preliminary classification result.

[0121] In this embodiment, the type classification module divides the graph data stream into different types based on a comprehensive judgment score and a preset judgment rule. The design of the judgment rule needs to balance the accuracy of classification and the resource utilization efficiency of the system.

[0122] Set classification boundary thresholds. The simplest judgment rule is a single threshold division. For example, set the classification boundary to 0.6. Data streams with a comprehensive judgment score greater than 0.6 are classified as "large flow - continuous graph traversal operation" type, and those less than or equal to 0.6 are classified as "small flow - short-term control signaling" type. This binary classification method is simple and direct, but it may have the problem of ambiguous boundaries.

[0123] For more refined classification, multiple thresholds can be set to divide data streams into multiple types. For example, a high threshold of 0.7 and a low threshold of 0.4 can be set. Data streams with a comprehensive judgment score greater than 0.7 are classified as "definitely large flow," those less than 0.4 as "definitely small flow," and those between 0.4 and 0.7 as "medium flow." For "definitely large flow," an OCS path is directly assigned; for "definitely small flow," an EPS path is directly assigned; and for "medium flow," more complex decision-making strategies can be adopted, such as further analyzing other auxiliary characteristics or using a hybrid transmission method.

[0124] Consider dynamic adjustments to classification. During system operation, dynamically adjust classification boundaries based on real-time network resource conditions. For example, when OCS resources are sufficient, the classification boundary threshold can be lowered to allow more data streams to use OCS, improving overall transmission performance; when OCS resources are scarce, the classification boundary threshold can be raised to strictly limit the number of data streams using OCS, avoiding resource contention and configuration delays.

[0125] A fuzzy classification method is introduced. Instead of rigidly classifying data streams into a single type, the method calculates the membership degree of each data stream to each type. For example, if a data stream has a comprehensive judgment score of 0.65, its membership degree to the "large stream" type is calculated to be 0.7, and its membership degree to the "small stream" type is calculated to be 0.3. In subsequent decision-making, probabilistic selection or a hybrid strategy can be made based on the membership degree.

[0126] After applying the judgment rules, a preliminary classification result is obtained. The preliminary classification result includes the identification information of the data flow and the classification type label. For example, if the data flow ID is 12345, the classification type is "large flow - persistent graph traversal operation".

[0127] S2.5: Add confidence information to the preliminary classification results and output the graph data stream classification results.

[0128] In this embodiment, the confidence calculation module adds confidence information to the classification results, quantifying the reliability of the classification judgment. Confidence reflects the certainty of the classification decision, providing a basis for subsequent exchange path selection and risk control.

[0129] The confidence level is calculated based on the distance between the comprehensive decision score and the classification boundary. If the comprehensive decision score is far from the classification boundary, it indicates that the data flow features are obvious, the classification certainty is high, and the confidence level is high. If the comprehensive decision score is close to the classification boundary, it indicates that the data flow features are ambiguous, the classification uncertainty is high, and the confidence level is low. For example, if the classification boundary is 0.6, and a data flow has a comprehensive decision score of 0.85, which is 0.25 from the boundary, the confidence level can be set to 0.9; another data flow has a comprehensive decision score of 0.62, which is only 0.02 from the boundary, and the confidence level can be set to 0.55.

[0130] Confidence scores can be calculated using linear mapping or nonlinear functions. In linear mapping, the confidence score equals the distance between the composite decision score and the classification boundary divided by the maximum possible distance, then mapped to the interval between 0.5 and 1. In nonlinear function methods, exponential functions, logarithmic functions, etc., can be used, making the confidence score increase faster the further away from the boundary.

[0131] The confidence level is calculated by combining historical classification accuracy statistics. The consistency between the actual transmission performance and classification prediction of data streams with historical composite scores within a specific range is analyzed. For example, if statistics show that 95% of data streams with composite scores between 0.8 and 0.9 exhibit large flow characteristics and perform well using OCS, then for the current data stream with a composite score of 0.85, the classification confidence level can be set to 0.95.

[0132] Consider the completeness and quality of the feature data. If certain features of a data stream cannot be accurately extracted, or if feature values ​​are abnormal, reduce the classification confidence. For example, if the graph traversal depth is missing due to packet encryption, the overall classification confidence should be reduced accordingly.

[0133] Confidence information is added to the classification results to form a complete graph data stream classification result. The classification result contains three parts: data stream identifier, classification type label, and confidence score. For example, the data stream ID is 12345, the classification type is "large flow - persistent graph traversal operation", and the confidence score is 0.88. The output classification results serve as important inputs for subsequent exchange path decisions. High-confidence classification results can directly apply the corresponding exchange strategy, while low-confidence classification results require more careful decision-making or further verification.

[0134] Example 4

[0135] This embodiment is a further refinement of S3 in Embodiment 1, such as... Figure 4 As shown, based on the graph data stream classification results, an optical circuit switching (OCS) path or an electrical packet switching (EPS) path is selected as the actual switching path for hybrid switching of the graph data streams using a switching path decision algorithm. This includes: monitoring the OCS and EPS resource status based on the graph data stream classification results and outputting a network resource status map; and, according to the network resource status map, directing the first graph data stream to the OCS optical path and the second graph data stream to the EPS channel, wherein the data volume of the first graph data stream is greater than the data volume of the second graph data stream.

[0136] S3.1: Based on the graph data stream classification results, monitor the status of OCS and EPS resources and output a network resource status graph.

[0137] In this embodiment, the resource monitoring module monitors the resource status of OCS and EPS in the hybrid switching network in real time, providing accurate resource information for switching path decisions. Resource monitoring covers multiple aspects, including resource availability, performance indicators, and configuration status.

[0138] Monitor OCS resource status. The OCS system is based on optical switching technology, establishing end-to-end optical path connections through an optical switching matrix. Monitor the number of available wavelengths in the OCS. Each fiber link can carry multiple wavelength channels, and each wavelength can establish an independent optical path. Count the number of currently idle wavelengths to reflect the available capacity of the OCS. Monitor the status information of established optical paths, including the source node, destination node, wavelength used, allocated bandwidth, establishment time, and estimated duration, to understand the occupancy of OCS resources.

[0139] Monitor the performance metrics of the OCS. Measure the actual transmission bandwidth of the optical path to verify whether the expected high bandwidth performance is achieved; measure the end-to-end delay of the optical path, including optical signal propagation delay and switching equipment processing delay, to ensure that low latency requirements are met; monitor the bit error rate and packet loss rate of the optical path to evaluate transmission quality.

[0140] Monitor the configuration overhead of the OCS. Record the time overhead of optical path establishment, from sending the configuration command to the actual completion of optical path establishment, which is usually in the range of milliseconds to seconds; record the time overhead of optical path teardown; count the number and frequency of optical switching matrix configurations to evaluate the configuration load.

[0141] Monitor EPS resource status. The EPS system is based on packet switching technology and forwards data packets through routers and switches. Monitor the bandwidth utilization of EPS links, calculate the ratio of current traffic to link capacity for each link, and identify congested and idle links. Monitor the queue status of EPS devices, including queue length, queue occupancy, and packet loss, reflecting the load and congestion level of the devices.

[0142] Monitor EPS performance metrics. Measure end-to-end packet delay, including propagation delay, queuing delay, and processing delay; measure delay jitter, i.e., the magnitude of delay variation; and calculate packet loss rate to assess network congestion and transmission reliability.

[0143] Monitor network topology and connectivity status. Obtain network topology information, including nodes, links, and connections; monitor link availability, identify faulty and backup links; monitor the configuration status of routing and forwarding tables.

[0144] All monitored resource status information is integrated to construct a network resource status diagram. The network resource status diagram is represented by a graph structure, where nodes represent network devices such as OCS switches, EPS routers, and servers, edges represent link connections, and corresponding status information such as available bandwidth, latency, and utilization is labeled on nodes and edges. The network resource status diagram provides a global view of the network, offering a comprehensive information foundation for switching path decisions.

[0145] S3.2: According to the network resource status diagram, the data stream of the first diagram is directed to the OCS optical path and the data stream of the second diagram is directed to the EPS channel, wherein the data volume of the data stream of the first diagram is greater than the data volume of the data stream of the second diagram.

[0146] In this embodiment, the path decision module intelligently selects the optimal switching path based on the graph data flow classification results and the network resource status graph. Path decision-making needs to comprehensively consider multiple factors such as data flow characteristics, resource availability, performance requirements, and configuration overhead.

[0147] For the first graph data stream classified as "large-flow persistent graph traversal operation" with high confidence (large flow, i.e., large data volume), the OCS optical path is preferentially selected for transmission. It is determined whether the OCS has available resources, whether there are idle wavelengths, and whether an optical path connection can be established between the source and destination nodes. If the OCS resources are available, the optimal optical path configuration scheme is calculated.

[0148] Select an appropriate wavelength channel. If multiple idle wavelengths are available, choose the optimal wavelength based on factors such as wavelength performance characteristics, current load, and historical usage. For example, prioritize wavelengths with stable performance and low bit error rate; or choose wavelengths with lighter current load that can provide higher bandwidth.

[0149] Determine the optical path routing. In the network topology, there may be multiple physical paths from the source node to the destination node, and the optimal path needs to be selected. Path selection can be based on the shortest path algorithm, such as Dijkstra's algorithm, using hop count or physical distance as a metric; or it can be based on resource optimization algorithms, comprehensively considering factors such as available bandwidth, latency, and reliability along the path to select the path with the best overall performance.

[0150] Estimate the optical path establishment time. Based on the configuration performance and current load of the OCS equipment, estimate the time required from sending the configuration command to the actual completion of optical path establishment. If the estimated establishment time is long, it may affect the timeliness of data stream transmission, which needs to be weighed in the decision-making process. For example, for latency-sensitive data streams, if the optical path establishment time exceeds the acceptable range, it may be necessary to consider transmitting via EPS first, and then switching after the optical path is established.

[0151] Generate OCS configuration instructions. These instructions include: port connection relationships of the optical switching matrix, specifying the mapping between input and output ports; wavelength allocation information, specifying the wavelength channels used; bandwidth parameters, specifying the bandwidth allocated to the optical path; and lifetime parameters, specifying the expected duration of the optical path or the conditions triggering its removal. The configuration instructions are then sent to the OCS controller, which drives the optical switching equipment to perform the configuration operation and establish the optical path connection.

[0152] For the second graph data stream classified as "small flow - short-time control signaling" (small flow, i.e., small data volume), the EPS channel is selected for transmission. EPS transmission does not require the pre-establishment of a dedicated connection, and the data packets are forwarded immediately upon arrival, resulting in a fast response speed, making it suitable for small-scale, bursty data transmission.

[0153] Optimize the forwarding path of EPS. Based on the network resource state diagram, calculate the optimal forwarding path from the source node to the destination node. Path calculation can use the shortest path algorithm or the load balancing algorithm to avoid congested links and select the less loaded path to improve transmission performance.

[0154] Configure Quality of Service (QoS) parameters. Based on data flow priority and performance requirements, configure QoS parameters. For high-priority control signaling, set a higher forwarding priority to process it first in the router queue, reducing queuing delays; allocate a certain amount of bandwidth to avoid packet loss due to network congestion; set latency and jitter limits to ensure real-time requirements are met.

[0155] Generate EPS configuration commands. Configuration commands include routing table entries specifying the destination address and next-hop information; forwarding rules specifying matching conditions and forwarding actions, such as flow classification rules based on five-tuples; and QoS policies specifying priority, bandwidth limits, queue scheduling algorithms, etc. Send the configuration commands to the relevant routers and switches to update their forwarding tables and policy configurations.

[0156] For data streams with low classification confidence or those on the classification boundary, a hybrid decision-making strategy is adopted. Initial transmission can be performed via EPS (Electronic Data Transfer) while simultaneously initiating the OCS (Optical Channel Response) optical path establishment process. Before the optical path is established, data is transmitted via EPS to avoid delays caused by waiting for optical path establishment; after the optical path is established, subsequent data transmission is switched to OCS to fully utilize the high-performance advantages of OCS. This strategy balances the needs for transmission timeliness and performance optimization.

[0157] A parallel transmission strategy can also be employed, using both OCS and EPS for transmission simultaneously. The data stream is divided into multiple sub-streams, with some sub-streams transmitted via OCS and others via EPS. The traffic allocation ratio between the two paths is dynamically adjusted based on real-time performance feedback. This strategy improves transmission reliability and flexibility but increases system complexity.

[0158] After path selection and configuration are completed, the data stream begins actual transmission through the selected switching path. The first graph data stream is transmitted through the established OCS optical path, enjoying a dedicated channel with high bandwidth and low latency for transmitting large-scale graph data; the second graph data stream is transmitted through the EPS channel, utilizing the flexibility of packet switching to quickly transmit small-scale control signaling. The hybrid switching architecture fully leverages the advantages of both OCS and EPS, achieving efficient transmission of graph data streams.

[0159] Example 5

[0160] This embodiment is a further refinement of S4 in Embodiment 1, such as... Figure 5As shown, performance data for evaluating the actual switching effect is collected, and machine learning optimization is performed on the classification threshold and the decision rule based on the performance data to update classification parameters and decision rules, which comprises: collecting in real time the performance data after actual switching of the graph data stream, and determining an evaluation index result; establishing and training a machine learning model, inputting the evaluation index result into the machine learning model, and outputting an optimized parameter set; and updating the switching decision rule based on the optimized parameter set, and generating the updated classification parameters and decision rule.

[0161] S4.1: collecting in real time the performance data after actual switching of the graph data stream, and determining an evaluation index result.

[0162] In this embodiment, after the graph data stream completes transmission, the performance data acquisition module collects comprehensive performance data for evaluating the effect of switching decisions. The performance data covers multiple aspects such as transmission efficiency, service quality and resource utilization.

[0163] Collecting transmission throughput data. The actual throughput of the data stream in the transmission process is counted, that is, the amount of data successfully transmitted per unit time. The average throughput is calculated to reflect the overall transmission efficiency; the time sequence of the throughput is recorded, the rate change in the transmission process is analyzed, and bottleneck stages are identified; and the actual throughput is compared with the theoretical maximum throughput to calculate the throughput utilization.

[0164] Collecting end-to-end delay data. The total time for data to be sent from a source end to received by a destination end is measured, including all parts of propagation delay, queuing delay, processing delay, configuration delay and the like. The delay of each data packet is recorded, and the average delay, maximum delay and minimum delay are counted; the standard deviation or jitter of the delay is calculated to reflect the stability of the delay; and the distribution characteristics of the delay are analyzed to identify abnormal delay events.

[0165] Collecting packet loss rate data. The number of data packets lost in the transmission process is counted, and the packet loss rate is calculated, that is, the number of lost packets divided by the total number of sent packets. The causes of packet loss are analyzed, such as queue overflow caused by network congestion, transmission interruption caused by link failure, packet damage caused by bit error and the like; and the time and position where packet loss occurs are recorded to identify problem links.

[0166] Collecting resource utilization data. For a data stream transmitted by OCS, the bandwidth utilization of an optical path is counted, that is, the ratio of an actual transmission rate to the capacity of the optical path; the occupation time of the optical path is recorded, and the time utilization of optical path resources is calculated. For a data stream transmitted by EPS, the bandwidth utilization of a related link is counted, and the influence on the load of an EPS network is analyzed.

[0167] Collect configuration overhead data. For OCS transmission, record the time overhead for optical path establishment, from sending the configuration command to the actual availability of the optical path; record the time overhead for optical path teardown; and count the number and frequency of configuration operations. For EPS transmission, record the time overhead for routing table updates and QoS policy configuration.

[0168] Collect transmission reliability data. Record error events during transmission, such as connection interruption, timeout retransmission, and path switching; calculate the transmission success rate, i.e., the proportion of data streams that successfully complete transmission out of the total data stream; and analyze the causes and impacts of transmission failures.

[0169] Based on the collected performance data, the evaluation metrics are calculated. A comprehensive performance score is defined, which normalizes and weights multiple performance metrics. For example, the throughput metric is normalized to the actual throughput divided by the theoretical maximum throughput, with a score range of 0 to 1; the latency metric is normalized to 1 minus the actual latency divided by the maximum acceptable latency, with lower latency resulting in a higher score; and the packet loss rate metric is normalized to 1 minus the packet loss rate, with no packet loss resulting in a score of 1.

[0170] Assign weights to each indicator, such as throughput (0.4), latency (0.3), packet loss rate (0.2), and resource utilization (0.1), and calculate the overall performance score. The overall performance score reflects the overall effectiveness of the switching decision; a high score indicates an excellent decision, while a low score indicates that the decision needs improvement.

[0171] Performance data is correlated with classification results. For each data stream, its classification type, confidence level, selected exchange path, and actual performance are recorded, forming a complete decision-effect data record. These records serve as training data for machine learning, used to optimize classification parameters and decision rules.

[0172] S4.2: Establish and train a machine learning model, input the evaluation index results into the machine learning model, and output an optimized parameter set.

[0173] In this embodiment, the machine learning optimization module establishes an intelligent learning model, which learns the optimal classification parameters and decision rules by analyzing historical decision and performance data. The construction and training process of the machine learning model includes steps such as data preparation, model selection, training optimization, and validation evaluation.

[0174] Prepare the training dataset. Collect complete records of the historical graph data stream, with each record including: a feature profile of the data stream, i.e., a multi-dimensional feature vector; classification decision information, including classification type, confidence level, and selected exchange path; actual performance data, including metrics such as throughput, latency, and packet loss rate; and a comprehensive performance score, which serves as the target label for learning.

[0175] Preprocess the training data. Handle missing values ​​by using mean imputation, median imputation, or interpolation based on similar records for records with missing features. Handle outliers by identifying and processing data that deviates significantly from the normal range to avoid interference from outliers in model training. Standardize the data by standardizing feature values ​​of different dimensions to a uniform scale, such as normalizing to the 0-1 interval or standardizing to a distribution with a mean of 0 and a standard deviation of 1.

[0176] Choose an appropriate machine learning algorithm. Depending on the characteristics of the problem and the scale of the data, various algorithms can be selected. For classification problems, decision trees, random forests, support vector machines, neural networks, etc., can be chosen; for regression problems, linear regression, ridge regression, gradient boosting regression trees, etc., can be chosen; for parameter optimization problems, Bayesian optimization, genetic algorithms, reinforcement learning, etc., can be chosen.

[0177] Taking the random forest algorithm as an example, this paper details the model construction and training process. Random forest is an ensemble learning method that constructs multiple decision trees and combines their prediction results. It has strong generalization ability and anti-overfitting ability, and is suitable for handling multidimensional features and nonlinear relationships.

[0178] Construct a random forest model. Set the model's hyperparameters, including the number of decision trees (e.g., 100 trees; more trees make the model more stable but increase computational cost); the maximum depth of each tree (e.g., 10 layers) to control tree complexity and avoid overfitting; the feature sampling ratio (e.g., 0.8, randomly selecting 80% of the features during training for each tree to increase the diversity between trees); and the sample sampling ratio (e.g., 0.8, randomly selecting 80% of the samples during training for each tree using a bootstrap sampling method).

[0179] Train the random forest model. Input the training dataset into the model, and the model automatically performs the following process: For each decision tree, randomly sample a subset of samples from the training data using sampling with replacement; randomly select a subset of features from the feature set; based on the sample subset and feature subset, construct the decision tree, recursively selecting the optimal splitting feature and split point to maximize the purity of the child nodes, which can be measured by the Gini coefficient or information entropy; repeat the above process to construct all decision trees.

[0180] After training, the random forest model contains 100 independent decision trees, each learning different patterns and rules from the data. For a new data stream, its features are input into the model, each tree provides a prediction, and the random forest combines the predictions of all trees, using either voting or averaging to obtain the final prediction.

[0181] During model training, model parameters are optimized to maximize prediction accuracy. Cross-validation is used to evaluate model performance, dividing the training data into K folds (e.g., 5 folds). Each time, 4 folds are used as the training set and 1 fold as the validation set, alternating between the two, and the average performance metric is calculated. Based on the validation results, model hyperparameters are adjusted, such as increasing the number of trees, adjusting tree depth, or changing the feature sampling ratio, to find the optimal parameter combination.

[0182] A well-trained machine learning model can predict the optimal swapping method and expected performance based on data flow characteristics. By inputting the current classification parameters and decision rules into the model, it analyzes their impact on system performance and outputs optimization suggestions.

[0183] Output the optimized parameter set. The optimized parameter set includes the adjusted feature thresholds. For example, if the model finds that when the current traffic size threshold T1 is set to 100MB, some data streams that are actually large are misclassified as small, it is recommended to adjust T1 to 120MB; or if the model finds that the duration threshold T2 is set to 5 seconds, which is too lenient, it is recommended to adjust it to 8 seconds to more accurately distinguish continuous transmission.

[0184] The optimized parameter set also includes adjusted feature weights. Model analysis revealed that, in the current network environment, graph traversal depth features contribute significantly to classification accuracy, and it is recommended to increase their weight from 0.2 to 0.3; while packet interval features have weak discriminative power, and it is recommended to reduce their weight.

[0185] Optimizing the parameter set may also include new decision rules. The model has discovered specific correlation patterns between certain feature combinations and performance. For example, when the traffic size is greater than 150MB and the graph traversal depth is greater than 6 hops, the performance improvement of using OCS for transmission is particularly significant. It is recommended to set specific decision rules for such feature combinations and prioritize the allocation of OCS resources.

[0186] S4.3: Based on the optimized parameter set, update the exchange decision rules and generate the updated classification parameters and judgment rules.

[0187] In this embodiment, the parameter update module updates the parameter configuration and decision rules of the classification system based on the optimized parameter set output by the machine learning model, thereby achieving adaptive optimization of the system. Parameter updates need to be performed cautiously to avoid frequent and drastic adjustments that could lead to system instability.

[0188] Evaluate the rationality of the optimized parameter set. Compare the optimized parameters with the current parameters to analyze the magnitude and direction of the adjustment. If the adjustment is too large, such as a threshold change exceeding 50%, further verification of its rationality is needed to avoid erroneous optimization due to data anomalies or model overfitting. Simulation testing can be used to apply the optimized parameters to historical data to evaluate their impact on classification accuracy and system performance, confirming the optimization effect.

[0189] A gradual update strategy is adopted. Instead of adjusting parameters to the optimized value all at once, adjustments are made gradually in multiple steps. For example, if the current traffic size threshold is 100MB and the optimization suggestion is 120MB, it can be adjusted to 110MB first, and the effect observed over a period of time. If performance is indeed improved, then it can be further adjusted to 120MB. Gradual updates reduce the risk of parameter adjustments and ensure system stability.

[0190] Update the feature threshold parameters. Update the optimized thresholds for flow size (T1), duration (T2), and graph traversal depth (T3) to the classification system configuration. After the update, the classification algorithm uses the new thresholds to make judgments, enabling it to more accurately identify data flow types.

[0191] Update the feature weight parameters. Update the configuration of the scoring calculation module with the optimized feature weights. After the update, the calculation of the comprehensive judgment score reflects the latest importance of each feature, and the classification result is more in line with actual performance requirements.

[0192] Update decision rules. Modify or add decision rules based on new rule suggestions in the optimization parameter set. For example, add specialized rules for specific feature combinations so that when the data stream meets specific conditions, the corresponding exchange strategy is applied directly without going through the conventional scoring and judgment process, improving decision efficiency and accuracy.

[0193] Update the classification boundary threshold. Based on the optimization results, adjust the comprehensive judgment score boundary that distinguishes different types of data streams. For example, adjust the classification boundary between "large stream" and "small stream" from 0.6 to 0.65 to make the classification more stringent or lenient, in order to adapt to the current network resource conditions and performance goals.

[0194] Record parameter update history. Save information such as the time, content, reason, and expected effect of each parameter update to establish a historical record of parameter evolution. The historical record helps to analyze the long-term trend of parameter adjustments, evaluate the effectiveness of optimization strategies, and provide a basis for subsequent rollback operations. If an update causes a performance degradation, it is possible to quickly roll back to the previous parameter configuration.

[0195] The system generates updated classification parameters and decision rule sets. These updated parameters and rules are integrated into the classification system, which uses the new configuration for data stream classification and exchange decisions in subsequent runs. Through a closed-loop process of continuous performance monitoring, machine learning optimization, and parameter updates, the system continuously learns and improves, resulting in consistently enhanced classification accuracy and exchange performance, achieving adaptive intelligent optimization.

[0196] Example 6

[0197] This embodiment is a further refinement of S5 in Embodiment 1, such as... Figure 6As shown, based on the updated classification parameters and judgment rules, dynamic matching of graph data streams with network resources is performed to achieve intelligent transmission of graph data streams. This includes: formulating a graph data stream processing strategy based on the updated classification parameters and judgment rules, wherein resources are reserved for high-priority graph data streams in the graph data stream processing strategy; adjusting the switching path configuration in real time based on the graph data stream processing strategy, performing batch configuration and pre-configuration, and outputting optimized configuration instructions; and performing dynamic matching of graph data streams with network resources based on the optimized configuration instructions to achieve intelligent transmission of graph data streams.

[0198] S5.1: Based on the updated classification parameters and judgment rules, formulate a graph data stream processing strategy, in which resources are reserved for high-priority graph data streams.

[0199] In this embodiment, the strategy formulation module formulates differentiated graph data stream processing strategies based on the optimized classification system to achieve reasonable allocation and efficient utilization of resources. The processing strategies cover multiple aspects such as priority management, resource reservation, and cache scheduling.

[0200] Define data flow priorities. Assign priority levels to data flows based on factors such as classification type, confidence level, and business importance. For example, divide data flows into high, medium, and low priorities (or first, second, and third priorities). High-priority data flows include: data flows classified as "large flow - persistent graph traversal operations" with a confidence level higher than 0.8. These data flows have high performance requirements and large transmission volumes, and should be prioritized; critical business data flows, such as real-time graph analysis and online transaction processing, have strict requirements for latency and reliability. Medium-priority data flows include: data flows with medium classification confidence levels, or general business data flows. Low-priority data flows include: data flows classified as "small flow - short-term control signaling," or non-critical background task data flows.

[0201] Reserve resources for high-priority data streams. In the OCS system, a certain number of wavelength channels are reserved specifically for high-priority data streams. For example, if the OCS system has a total of 10 available wavelengths, 3 wavelengths can be reserved as dedicated high-priority channels to ensure that high-priority data streams can immediately obtain OCS resources upon arrival, without waiting or contention. The number of reserved resources is dynamically adjusted based on historical statistics and business forecasts, increasing reservations during peak business periods and decreasing reservations during off-peak periods to improve resource utilization.

[0202] In the EPS system, bandwidth is reserved for high-priority data flows. Through traffic engineering and QoS mechanisms, a certain percentage of bandwidth is reserved on critical links, for example, 20% of the bandwidth is reserved for high-priority traffic. High-priority packets enjoy priority forwarding rights in router queues, and strict priority scheduling or weighted fair queue scheduling is used to ensure that their latency and packet loss rate meet the requirements.

[0203] Design a caching strategy. For data streams that cannot be transmitted immediately, manage the cache according to priority. High-priority data streams are placed at the front of the cache queue and have priority for transmission; low-priority data streams may have to wait for a longer time, or even be dropped or delayed when resources are scarce. Since cache capacity is limited, a reasonable cache replacement strategy needs to be designed, such as Least Recently Used (LRU) or First-In-First-Out (FIFO), to determine which data streams are removed from the cache when it is full.

[0204] Develop a traffic shaping strategy. For bursty data streams, employ traffic shaping techniques to smooth their transmission rate and prevent network congestion caused by excessive instantaneous traffic. Traffic shaping can use algorithms such as the token bucket or leaky bucket to limit the peak and average rates of the data stream, ensuring they conform to the network's carrying capacity.

[0205] Develop a load balancing strategy. When multiple paths are available, dynamically allocate data flows based on the load status of each path to prevent some paths from being overloaded while others are idle. Load balancing can be based on flow granularity, distributing different data flows to different paths; or it can be based on packet granularity, distributing different packets of the same data flow across multiple paths, but attention must be paid to packet out-of-order issues.

[0206] Develop fault recovery strategies. Pre-plan backup paths and recovery mechanisms so that when the primary path fails, a rapid switch to the backup path ensures business continuity. For high-priority data streams, simultaneous transmission via primary and backup paths can improve reliability, although this consumes more resources, it ensures that critical business operations are not affected.

[0207] S5.2: Based on the graph data stream processing strategy, adjust the switching path configuration in real time, perform batch configuration and pre-configuration, and output optimization configuration instructions.

[0208] In this embodiment, the configuration optimization module dynamically adjusts the switching path configuration according to the graph data stream processing strategy, optimizes the configuration process, reduces configuration overhead, and improves system response speed. Configuration optimization includes techniques such as real-time adjustment, batch configuration, and pre-configuration.

[0209] Real-time monitoring of data stream arrivals and network status changes. When a new data stream arrives, it is immediately classified and determined, and the switching path is decided based on the classification result and the current resource status. If the network status changes, such as link congestion, equipment failure, or resource release, the impact on existing data streams is promptly assessed, and path adjustments are made as necessary.

[0210] Real-time adjustment of switching path configuration. For data streams already in transmission, if their actual transmission characteristics are detected to be inconsistent with classification predictions, dynamic adjustments are made. For example, if a data stream is classified as a "small stream" and transmitted using EPS, but during actual transmission it is found that its data volume far exceeds expectations and the duration is very long, it is switched to OCS transmission to obtain better performance. Path switching requires coordination between the source and destination ends to ensure data continuity and consistency, and to avoid data loss or duplication.

[0211] Perform batch configuration. Addressing the millisecond-level configuration overhead of OCS, multiple data streams requiring optical path establishment are aggregated and configured uniformly. For example, within a configuration cycle, all data stream requests needing to establish an OCS optical path are collected, and their source and destination addresses are analyzed. If multiple data streams have the same source and destination node pair, they can share the same optical path, transmitting multiple data streams on a single optical path using wavelength division multiplexing (WDM) or time division multiplexing (TDM) techniques. Batch configuration reduces the number of configuration operations, lowers overall configuration overhead, and improves the utilization efficiency of OCS resources.

[0212] Perform pre-configuration. Based on a data flow prediction model, predict potential future data flows and pre-establish optical path connections. The prediction model can analyze the arrival time patterns and periodicities of data flows based on historical statistics to predict the type and quantity of data flows that may arrive within a future time window. For example, if it is found that 9:00 AM to 11:00 AM is the peak period for graph computing tasks, and a large number of "large-flow-persistent graph traversal operation" type data flows arrive, then some OCS optical paths will be pre-established starting at 8:50 AM, so that the data flows can be used immediately upon arrival, eliminating configuration waiting time.

[0213] Pre-configuration requires a balance between resource consumption and response speed. Pre-established optical paths, if unused, result in resource waste; however, without pre-configuration, data streams must wait for optical paths to be established upon arrival, impacting transmission timeliness. A probabilistic pre-configuration strategy can be adopted, determining the number of pre-configured optical paths based on the predicted arrival probability of the data stream—more paths for high-probability arrivals, and fewer or no paths for low-probability arrivals.

[0214] Optimize the generation and issuance of configuration commands. Translate configuration decisions into specific device configuration commands, the format of which must conform to the device's control interface specifications. For OCS devices, configuration commands may use OpenFlow, GMPLS, or proprietary protocols from the device vendor; for EPS devices, configuration commands may use NETCONF, RESTCONF, or CLI commands.

[0215] Batch generation of configuration commands combines multiple configuration operations into one or a few commands, reducing control communication overhead. For example, OpenFlow's batch flow table distribution feature can be used to distribute multiple flow table entries at once, instead of distributing them one by one.

[0216] Optimize the order and timing of configuration command issuance. For interdependent configuration operations, determine the issuance order according to the dependency relationship to ensure the correctness of the configuration. For example, when establishing an end-to-end optical path, the switching matrix of the intermediate node needs to be configured first, and then the access port of the edge node needs to be configured. For time-sensitive configurations, issue them at the last moment before the data stream arrives to reduce the time of pre-occupancy of resources.

[0217] Output an optimized configuration command set. This set contains all necessary configuration operations, including establishing, adjusting, and tearing down OCS optical paths, updating EPS routing tables, and configuring QoS policies. The configuration commands are optimized and orchestrated to ensure efficient execution, minimize configuration overhead, and maximize system performance.

[0218] S5.3: Based on the optimized configuration instructions, perform dynamic matching of graph data streams and network resources to achieve intelligent transmission of graph data streams.

[0219] In this embodiment, the resource matching execution module coordinates the transmission requirements of the data stream with the supply of network resources according to the optimization configuration instructions, achieving dynamic matching and efficient transmission. Resource matching is a continuous and dynamic process that requires real-time response to changes to maintain the optimal state of the system.

[0220] Execute configuration commands to establish or adjust switching paths. Optimization configuration commands are sent to the corresponding network devices, which then perform configuration operations and update their internal states. For OCS devices, the optical switching matrix is ​​driven, specified port connections are established, wavelength resources are allocated, and after the optical path is established, link testing is performed to verify the connectivity and performance indicators of the optical path, ensuring that transmission requirements are met. For EPS devices, the routing table and forwarding table are updated, QoS policies are configured, and the devices begin forwarding data packets according to the new configuration.

[0221] Monitor the configuration execution results. Check whether the configuration operation was completed successfully and whether the expected results were achieved. If the configuration fails, such as optical path establishment failure or routing configuration conflicts, promptly diagnose and recover from the fault, try alternative solutions, or roll back to the previous configuration state. Record the configuration execution log, including configuration time, configuration content, execution results, error messages, etc., for subsequent analysis and optimization.

[0222] Initiate data stream transmission. After configuration, the source end is notified to begin data transmission, and the data stream is transmitted through the established switching path. For OCS transmission, data is transmitted as optical signals on a dedicated optical path, enjoying high-bandwidth, low-latency, and congestion-free high-quality service. For EPS transmission, data packets are forwarded hop-by-hop by routers, with priority scheduling and bandwidth allocation based on the configured QoS policy.

[0223] Real-time monitoring of the transmission process. Continuously monitor the data stream's transmission status, including current transmission rate, amount of data transmitted, amount of data remaining, and estimated completion time. Monitor network resource usage, including link bandwidth utilization, device load, and queue length. Monitor transmission quality metrics, including latency, jitter, packet loss rate, and bit error rate.

[0224] Dynamically adjust resource allocation. Based on real-time monitoring results, dynamically adjust resource allocation strategies. If the transmission rate of a data stream is lower than expected, analyze the cause, which may be path congestion, insufficient resources, or improper configuration, and take corresponding measures, such as increasing bandwidth allocation, switching to an alternative path, or adjusting QoS parameters. If the utilization rate of a resource is too low, consider allocating it to other data streams or releasing the resource to save costs.

[0225] Handle data stream completion and resource release. Once the data stream transmission is complete, promptly release the occupied resources. For OCS optical paths, after the data stream ends, dismantle the optical path connection, releasing wavelength resources for other data streams. Optical path dismantling also incurs some time overhead, but it is usually shorter than the setup time. For EPS resources, delete the corresponding flow table entries and QoS configurations, releasing bandwidth reservations.

[0226] This system coordinates resource contention among multiple data streams. When multiple data streams simultaneously request the same resource, arbitration is conducted based on priority and fairness principles. Higher-priority data streams are given priority access to resources, while lower-priority data streams may need to wait or use suboptimal resources. A fair scheduling algorithm, such as a weighted fair queue (WFQ), is employed to guarantee priority while preventing low-priority data streams from experiencing prolonged starvation.

[0227] Achieve global optimization. From a holistic system perspective, coordinate the configuration of various data flows and resources to pursue optimal global performance, rather than merely achieving local optima for individual data flows. Employ centralized control or distributed collaborative mechanisms to collect global information and make joint optimization decisions. For example, when resources are scarce, the performance of some low-priority data flows may be sacrificed to ensure the needs of high-priority data flows, thereby maximizing overall business value.

[0228] Through the aforementioned dynamic matching and intelligent transmission process, the transmission needs of graph data streams and the supply of network resources are efficiently matched. The system fully leverages the advantages of both OCS and EPS while overcoming their respective limitations, achieving high-performance, high-reliability, and high-efficiency graph data transmission. The entire system forms a complete closed loop from feature extraction, intelligent classification, path decision-making, performance monitoring, machine learning optimization to dynamic resource matching, continuously learning and improving to enhance the system's intelligence level and transmission performance.

[0229] Example 7

[0230] This embodiment adds a new step to S5 in Embodiment 1. Before performing dynamic matching of graph data streams and network resources, the method further includes: acquiring historical transmission records of the graph data stream, extracting traffic change patterns and resource occupancy characteristics, and fusing them to obtain a graph data stream state feature set; based on the graph data stream state feature set, constructing a Markov state transition model, and applying a reachability analysis algorithm to calculate the state transition probability to obtain traffic state prediction results at each time point; based on the traffic state prediction results and the existing network resource status, calculating the resource demand within a future time window; based on the resource demand within the future time window, performing path planning in conjunction with resource constraints of OCS and EPS to obtain a preset resource allocation strategy; and based on the preset resource allocation strategy, collaboratively optimizing the resource demands of multiple data streams to achieve dynamic matching of the graph data stream and network resources.

[0231] Step A: Obtain the historical transmission records of the graph data stream, extract the traffic change patterns and resource usage characteristics, and fuse them to obtain the graph data stream state feature set.

[0232] In this embodiment, the historical data analysis module retrieves historical transmission records of graph data streams from system logs and the monitoring database. The historical records contain detailed information about all graph data streams over a past period, such as the most recent month or three months, including data stream identifier, arrival time, duration, traffic volume, transmission path, resource consumption, and performance metrics.

[0233] Extract patterns in traffic changes. Analyze the arrival time series of historical data streams to identify time patterns. Statistically analyze the arrival frequency of data streams in different time periods. For example, by hourly statistics, it is found that peak arrival times for data streams are 9:00 AM to 11:00 AM and 2:00 PM to 4:00 PM on weekdays; by weekday statistics, it is found that traffic on Mondays and Fridays is usually higher than during the week. Identify periodic patterns, such as daily, weekly, and monthly traffic change cycles. Some businesses may have obvious daily or weekly cycle characteristics.

[0234] Analyze trends in traffic volume. Statistically analyze average traffic volume, peak traffic, and traffic distribution over different time periods to identify trends of traffic growth or decline. For example, a recent increase in the proportion of large data streams may be due to business expansion or changes in application models.

[0235] Extract resource usage characteristics. Analyze the historical data stream's usage of OCS and EPS resources. Statistically analyze the usage frequency, average occupancy time, and peak concurrent optical path count of OCS optical paths to identify OCS resource load patterns. Statistically analyze the bandwidth usage, congestion frequency, and average latency of EPS links to identify bottlenecks in EPS resources.

[0236] Analyze the correlation between resource consumption and data stream characteristics. For example, it was found that data streams larger than 200MB typically occupy the OCS optical path for more than 30 seconds; data streams with a graph traversal depth greater than 8 hops typically require bandwidth exceeding 1Gbps. These correlations help predict the resource requirements of future data streams.

[0237] By integrating traffic variation patterns and resource consumption characteristics, a graph data stream state feature set is constructed. This state feature set is represented as a vector or matrix, containing information such as traffic variation patterns, traffic size distribution, and resource demand distribution over time. For example, a 24-dimensional time feature vector can be constructed, with each dimension corresponding to an hour of the day, and the vector value representing the average traffic or data stream arrival frequency for that hour.

[0238] Step B: Based on the graph data flow state feature set, construct a Markov state transition model and apply the reachability analysis algorithm to calculate the state transition probability, thereby obtaining the flow state prediction results at each time point.

[0239] In this embodiment, the state prediction module constructs a Markov state transition model based on the extracted state feature set to predict future flow states. The Markov model assumes that the future state of the system depends only on the current state and is independent of historical states, making it suitable for modeling time-dependent stochastic processes.

[0240] Define the flow state space. Divide the flow level of the graph data stream into several discrete states, for example, defining five states: very low flow, low flow, medium flow, high flow, and very high flow. The state division can be based on the statistical distribution of historical flow data, for example, dividing the flow according to the quintiles, with each state corresponding to 20% of the data.

[0241] Construct a state transition matrix. Analyze historical data to count the frequency of transitions from one state to another and calculate the state transition probability. For example, count the number of times in historical records the current state is in "medium flow" and the next state is in "high flow," divide this number by the total number of times the current state is in "medium flow" to obtain the transition probability. Construct a 5x5 state transition matrix, where the element in the i-th row and j-th column represents the probability of transitioning from state i to state j.

[0242] Reachability analysis algorithms are applied to calculate the probability of reaching each state after several transitions from the current state. Reachability analysis can be implemented using matrix exponentiation. For example, multiplying the current state vector by the state transition matrix yields the probability distribution of the state at the next time step; multiplying it again by the state transition matrix yields the probability distribution of the state at the time step after that, and so on.

[0243] Consider the non-stationarity of the time dimension. The transition probability of actual traffic states may change over time; for example, the transition patterns differ between weekdays and weekends, and between daytime and nighttime. A time-varying state transition matrix can be constructed, setting different transition probabilities for different time periods to improve prediction accuracy.

[0244] The predicted flow conditions at each time point are obtained. These predictions are expressed as probability distributions. For example, predicting that one hour from now, the probability of the system being in an "extremely low flow" state is 0.1, in a "low flow" state is 0.2, in a "medium flow" state is 0.4, in a "high flow" state is 0.25, and in an "extremely high flow" state is 0.05. Based on these probability distributions, the most likely flow conditions can be predicted, or the expected value and variance of the flow can be calculated, providing a basis for resource planning.

[0245] It should be noted that reachability analysis is a state-space analysis method based on graph theory and probability theory, used to calculate the probability of transitioning from one state to another in a Markov model. In graph data flow prediction scenarios, this algorithm analyzes historical state transition paths, identifies all possible state transition sequences, and calculates a transition probability matrix based on historical frequency statistics, thereby predicting the potential flow state of the graph data flow at future points in time. Reachability analysis considers not only direct one-step transitions but also the cumulative effect of multi-step transition paths, enabling it to capture indirect relationships between states and long-term evolutionary trends.

[0246] The core principle of reachability analysis algorithms is based on the state transition theory of Markov chains and the reachability theory of graphs. First, the different flow states of a graph data stream are abstracted as discrete state nodes in a Markov chain, such as low flow, medium flow, high flow, and burst flow states. The transition relationships between states form a directed graph, where edges represent the possibility of state transitions, and the edge weights represent the transition probabilities.

[0247] The first step in reachability analysis is to construct a state transition graph. By analyzing historical transmission records, we count the number of times a data stream in state i at time t transitions to state j at the next time t+1. Let N represent the number of transitions from state i to state j, and let N' represent the total number of transitions starting from state i. Then, the one-step transition probability P is equal to N divided by N'. This probability reflects the direct transition relationship between states.

[0248] The second step in reachability analysis is calculating multi-step transition probabilities. In practical applications, the state evolution of a data flow may involve multiple intermediate states. For example, transitioning from a low-flow state to a high-flow state might first involve a transition to a medium-flow state, and then to a high-flow state. Multi-step transition probabilities are calculated using matrix multiplication; the nth power of the n-step transition probability matrix P is equal to the nth power of the one-step transition probability matrix P. The element P in the matrix represents the probability of transitioning from state i to state j in n steps.

[0249] The third step in reachability analysis is identifying the set of reachable states. Given an initial state and a time window, the algorithm calculates all possible reachable states within that time window, along with the cumulative probability of reaching each state. Reachability is determined based on a transition probability threshold; if the transition probability from the initial state to the target state is greater than a preset threshold, the target state is considered reachable. This step helps the system identify high-probability future states and filter out low-probability noisy states.

[0250] The fourth step in reachability analysis is to consider the absorption and periodicity of states. Some states may exhibit absorption characteristics, meaning that once entered, it is difficult to transition to other states, such as a state with consistently high traffic volume. Some state transitions may exhibit periodicity, such as traffic peaks occurring at specific times of the day. The algorithm improves prediction accuracy by analyzing the eigenvalues ​​and eigenvectors of the transition matrix to identify steady-state distributions and periodic patterns.

[0251] The implementation of the reachability analysis algorithm consists of five stages: data preprocessing, state discretization, transition matrix construction, reachability calculation, and prediction output.

[0252] During the data preprocessing stage, the system extracts time-series data from historical transmission records, including features such as traffic volume, resource utilization, and transmission latency at each time point. Data cleaning removes outliers and missing values, such as negative traffic values ​​due to measurement errors or extreme values ​​exceeding physical limits. Data normalization maps features with different dimensions to a unified numerical range, such as the zero-to-one interval, facilitating subsequent state partitioning and probability calculations.

[0253] In the state discretization stage, the system maps continuous flow values ​​to a discrete state space. Common discretization methods include equal-width binning, equal-frequency binning, and cluster binning. Equal-width binning divides the flow value range evenly into several intervals, each interval corresponding to a state; for example, 0 to 50 Mbits per second is a low flow state, 50-100 Mbits per second is a medium flow state, and above 100 Mbits per second is a high flow state. Equal-frequency binning, based on the distribution of historical data, ensures that each state contains approximately the same number of samples, suitable for situations with uneven data distribution. Cluster binning uses K-means clustering or hierarchical clustering algorithms to automatically divide states based on the similarity of flow characteristics, revealing the inherent structure of the data. The choice of the number of states needs to balance the model's expressive power and computational complexity; too few states lead to information loss, while too many states result in inaccurate transition probability estimation. In practice, the number of states is usually set to three to ten.

[0254] During the transition matrix construction phase, the system statistically analyzes all adjacent time point state pairs in historical data to construct a state transition frequency matrix. It iterates through the time series, recording the current state *s* and the state *s* of the next time point at each time *t*, and incrementing the corresponding element in the transition count matrix by one. After iteration, each row is normalized so that the row sum equals 1, resulting in the transition probability matrix. To avoid the zero-probability problem—that is, some state transitions that have never appeared in historical data but may actually occur—Laplace smoothing is used, adding a small smoothing constant, such as 0.1, to each transition count before re-normalizing. The construction of the transition matrix also needs to consider time dependence. If the state transition patterns of the data stream change over time, a time-varying transition matrix can be constructed, using different transition probabilities for different time periods, such as different matrices for weekdays and weekends, and different matrices for daytime and nighttime.

[0255] In the reachability calculation phase, the system calculates the reachability probabilities from the current state to each future state based on the transition matrix. Let the current time be t and the current state be s, predicting the state distribution after k future time steps. Initialize the state probability vector, with the probability of the current state being 1 and the probabilities of other states being zero. Iteratively calculate the state probability distribution for each future time step; the probability vector at step n is equal to the probability vector at step n-1 multiplied by the transition matrix. After k iterations, obtain the state probability distribution vector at future time t+k, where each element represents the probability that the data stream is in the corresponding state.

[0256] Reachability analysis also includes path enumeration and probability accumulation. For scenarios requiring detailed analysis, the system enumerates all possible paths from the current state to the target state and calculates the probability of each path. The path probability is equal to the product of all transition probabilities along the path. For example, the path probability from state A to state C via state B is equal to P*P. If multiple paths lead to the same target state, the total reachability probability of the target state is equal to the sum of all path probabilities. Path enumeration is suitable for short time windows and a small number of states. For long time windows and a large number of states, the number of paths grows exponentially, leading to excessive computational complexity. In such cases, matrix exponentiation is more efficient.

[0257] In the prediction output phase, based on the calculated state probability distribution, the predicted flow state results for each time point are output. The prediction results include the most likely state, the probability distribution, and the confidence interval. The most likely state is the state with the highest probability, representing the system's point prediction for the future state. The probability distribution provides complete uncertainty information, representing the likelihood of each state occurring. The confidence interval is calculated using cumulative probability; for example, a 95% confidence interval contains the set of states with a cumulative probability of 95%. The prediction results can also be converted back to continuous flow values, calculated using representative or expected values ​​for each state. For example, a low flow state corresponds to 25 Mbits per second, a medium flow state to 75 Mbits per second, and a high flow state to 150 Mbits per second. The predicted flow is equal to the weighted sum of the representative values ​​for each state multiplied by their corresponding probabilities.

[0258] Step C: Based on the traffic state prediction results and the existing network resource state, calculate the resource demand within the future time window.

[0259] In this embodiment, the resource demand calculation module calculates the resource demand within the future time window based on the traffic status prediction results and the existing network resource status, including OCS optical path demand and EPS bandwidth demand.

[0260] Traffic state predictions are translated into specific resource requirements. Based on historical statistics, a mapping relationship between traffic state and resource requirements is established. For example, when the system is in a "high traffic" state, the historical average requires 8 OCS optical paths and 500Mbps EPS bandwidth; when it is in a "medium traffic" state, the historical average requires 4 OCS optical paths and 300Mbps EPS bandwidth.

[0261] Based on the predicted probability distribution of traffic states, the expected value of resource demand is calculated. For example, if the probability distribution of the system being in each state one hour from now is known, then the expected value of OCS optical path demand is equal to the sum of the optical path demand in each state multiplied by the corresponding probability. Similarly, the expected value of EPS bandwidth demand is calculated.

[0262] Consider the uncertainty and volatility of resource demand. In addition to expected values, it's also necessary to consider the variance or confidence interval of demand. For example, calculate a 95% confidence upper bound for resource demand, meaning there's a 95% probability that actual demand will not exceed this value. During resource planning, reserves can be made based on this confidence upper bound to cope with demand fluctuations and avoid resource shortages.

[0263] Assess the resource supply and demand match based on the current network resource status. Compare the predicted resource demand with the current available resources to identify resource gaps or redundancies. For example, if it is predicted that 5 OCS optical paths will be needed in the next hour, but there are currently only 3 idle optical paths, there is a gap of 2 optical paths. It is necessary to release some occupied optical paths in advance or adjust the resource allocation strategy.

[0264] Calculate resource requirements for different time windows. This involves predicting not only the demand at the next moment, but also the demand at multiple future points in time, such as the resource demand in the next 1 hour, 2 hours, 4 hours, and 8 hours, forming a time series of resource demands. This time series of resource demands provides a basis for forward-looking resource planning, allowing for advance preparation and avoiding resource shortages or allocation delays caused by last-minute responses.

[0265] Step D: Based on the resource demand within the future time window, and in conjunction with the resource constraints of OCS and EPS, perform path planning to obtain a pre-set resource allocation strategy.

[0266] In this embodiment, the path planning module performs forward-looking path planning based on the predicted resource demand and the resource constraints of OCS and EPS, and formulates a pre-set resource allocation strategy to reserve and configure resources in advance for data streams that may arrive in the future.

[0267] Analyze the resource constraints of OCS and EPS. OCS constraints include: a limited number of available wavelengths, with each fiber link capable of carrying only a limited number of wavelength channels; the time overhead of establishing and tearing down optical paths, with frequent configurations impacting system performance; and the exclusive use of resources once an optical path is established, preventing other data streams from sharing them. EPS constraints include: limited link bandwidth, with total traffic not exceeding link capacity; limited device processing capacity, with excessive load leading to increased latency and packet loss; and while EPS resources can be dynamically shared, congestion and contention issues exist.

[0268] Develop a pre-configuration strategy for OCS resources. Based on the predicted OCS optical path demand, establish some optical paths in advance. Select the most likely source-destination node pairs to be used, and identify frequently accessed node pairs based on historical statistics and business patterns, and pre-establish optical paths for them. For example, if it is found that data flow from node A to node B occurs frequently and usually arrives around 9:00 AM, then pre-establish the optical path from A to B at 8:50 AM so that the data flow can be used immediately upon arrival.

[0269] The number of pre-configured optical paths needs to balance resource utilization and response speed. Too many pre-configured paths lead to resource waste; if these paths are not used, the occupied wavelength resources cannot be used by other data streams. Too few pre-configured paths, on the other hand, cannot effectively reduce configuration waiting time. A probabilistic pre-configuration strategy can be adopted based on predicted demand probabilities, pre-establishing optical paths with high-probability demand and establishing those with low-probability demand on demand.

[0270] Develop an EPS resource reservation strategy. Based on the predicted EPS bandwidth demand, reserve bandwidth on critical links. Utilize traffic engineering techniques, such as MPLS traffic engineering and SDN flow table configuration, to reserve transmission paths and bandwidth resources for predicted data flows. The reservation strategy needs to consider bandwidth allocability and sharing; EPS bandwidth can be dynamically shared among multiple data flows, but excessive reservation should be avoided to prevent resource waste.

[0271] Perform joint path planning. Taking into account the resource constraints and predicted demand of OCS and EPS, formulate a hybrid switching path planning scheme. For predicted high-volume data flows, plan the OCS optical path; for predicted low-volume data flows, plan the EPS path; for data flows with high uncertainty, plan a flexible hybrid strategy, initially using EPS and switching to OCS when necessary.

[0272] Consider the time dimension of path planning. Resource requirements vary at different times, necessitating dynamic adjustments to path planning. For example, during peak business periods, increase the number of pre-configured OCS optical paths to improve system capacity; during off-peak periods, reduce pre-configuration to free up resources and reduce overhead. Path planning generates a time-series resource allocation scheme to guide resource configuration at different points in time.

[0273] A pre-defined resource allocation strategy is obtained. This strategy includes a specific resource configuration plan, such as establishing optical path L1 at time T1, connecting nodes A and B, using wavelength W1, with an estimated duration of 30 minutes; and reserving EPS path P1 at time T2, with a bandwidth of 500Mbps, routed from node C to node D via node E. This pre-defined resource allocation strategy provides guidance for the system's proactive resource management, improving the system's responsiveness to future needs.

[0274] Step E: Based on the preset resource allocation strategy, the resource requirements of multiple data streams are collaboratively optimized to achieve dynamic matching between the graph data stream and network resources.

[0275] In this embodiment, the collaborative optimization module, based on a pre-defined resource allocation strategy, globally coordinates and optimizes the resource requirements of multiple graph data streams, achieving efficient resource utilization and overall performance improvement. Collaborative optimization requires comprehensive consideration of the mutual influence, resource competition, and cooperative relationships among multiple data streams.

[0276] Identify resource contention relationships between data flows. Contention exists when multiple data flows simultaneously request the same network resources. For example, multiple data flows may need to use wavelength resources on the same fiber optic link or pass through the same network bottleneck node. Analyze historical data and forecasts to identify potential time points and locations of resource contention, and develop coordination strategies in advance.

[0277] Identify opportunities for collaboration between data streams. Some data streams may have similar transmission paths or timing characteristics, allowing them to share resources or coordinate transmission. For example, multiple data streams with the same source and destination nodes can reuse the same OCS optical path, using time-division multiplexing or wavelength-division multiplexing techniques to transmit multiple data streams on a single optical path, improving optical path utilization. Alternatively, multiple data streams with staggered transmission times can sequentially use the same resources, avoiding simultaneous contention.

[0278] Establish a multi-stream collaborative optimization model. The resource allocation problem among multiple data streams is modeled as an optimization problem, with the objective of maximizing overall system performance. Constraints include resource capacity limitations, quality of service requirements, and fairness constraints. Optimization objectives can be maximizing total throughput, minimizing average latency, maximizing resource utilization, or a weighted combination of these objectives.

[0279] Optimization algorithms are employed to solve multi-stream resource allocation schemes. For small-scale problems, exact algorithms such as linear programming, integer programming, and dynamic programming can be used to find the optimal solution. For large-scale problems, the computational complexity of exact algorithms is too high; therefore, heuristic or approximate algorithms, such as greedy algorithms, genetic algorithms, and simulated annealing algorithms, can be used to find near-optimal solutions within a reasonable timeframe.

[0280] Implement a collaborative scheduling strategy. Based on the results of the optimization algorithm, formulate specific resource allocation and scheduling schemes. For high-priority data streams, allocate optimal resources first to ensure their performance requirements are met. For low-priority data streams, flexibly adjust resource allocation to make way for high-priority data streams, while ensuring basic quality of service.

[0281] Dynamically adjust resource allocation. During system operation, the actual arrival of data streams may deviate from predictions, necessitating dynamic adjustments to the resource allocation scheme. When actual demand exceeds predictions, activate emergency resources or adjust existing allocations, such as compressing resources for low-priority data streams to make room for newly arriving high-priority data streams. When actual demand is lower than predictions, release excess reserved resources to improve resource utilization efficiency.

[0282] Handling resource allocation conflicts. When the resource needs of multiple data streams cannot be met simultaneously, arbitration is conducted based on priority and fairness principles. A priority preemption mechanism can be used, allowing higher-priority data streams to preempt resources from lower-priority data streams. However, the subsequent transmission of preempted data streams must be properly handled to prevent them from experiencing prolonged starvation. A time-slice round-robin mechanism can also be used, allocating time slices to each data stream and allowing them to use resources in turn, ensuring fairness.

[0283] This enables resource coordination across time windows. It optimizes resource allocation not only at the current moment but also considers resource demands in future time windows. For example, if a resource is currently idle, but forecasts indicate that a high-priority data stream will soon require it, requests from low-priority data streams can be rejected to reserve the resource for future high-priority needs. This cross-time window coordination improves the foresight and overall optimization of resource allocation.

[0284] Monitor the effectiveness of collaborative optimization. Collect performance data after the implementation of collaborative optimization and evaluate its effects. Compare system performance indicators before and after optimization, such as increased total throughput, reduced average latency, and improved resource utilization, to verify the effectiveness of collaborative optimization. Analyze problems and shortcomings in the optimization process, such as the performance of certain data streams not meeting expectations or uneven resource allocation, to provide directions for improvement in future optimizations.

[0285] Through multi-stream collaborative optimization, the system achieves an improvement from single-stream local optimization to multi-stream global optimization, fully tapping the collaborative potential of resources and improving overall transmission efficiency. Combined with the aforementioned technologies such as traffic prediction, resource pre-provisioning, and path planning, the system possesses powerful forward-looking resource management capabilities, proactively adapting to future traffic changes and achieving efficient dynamic matching of graph data streams and network resources.

[0286] Example 8

[0287] This embodiment adds new steps to Embodiment 1. The method further includes: acquiring the transmission behavior of the graph data stream, collecting traffic changes, resource usage, and performance index data; based on the traffic changes, resource usage, and performance index data, applying time series analysis methods to extract behavioral patterns, and identifying the inter-flow influence relationships in the behavioral patterns through Granger causality tests to obtain a preliminary causal association network; using the preliminary causal association network to calculate the influence propagation delay and propagation intensity, and outputting a weighted time-series causal graph; identifying key influence paths and propagation chains in the weighted time-series causal graph to obtain a causal propagation model; and predicting the chain reaction of decision adjustment based on the causal propagation model to achieve optimal decision-making for multi-flow collaboration.

[0288] Step F: Obtain the transmission behavior of the graph data stream, and collect data on traffic changes, resource usage, and performance indicators.

[0289] In this embodiment, the behavioral data acquisition module continuously monitors the transmission behavior of the graph data stream, collecting multi-dimensional behavioral data to provide a data foundation for causal relationship analysis. The transmission behavioral data covers multiple aspects such as traffic characteristics, resource usage, and performance.

[0290] Collect traffic variation data. For each graph data stream, record its traffic variation curve over time. Sample the instantaneous transmission rate of the data stream at different time points, for example, sampling once per second, to obtain time series data of the traffic. Record abrupt traffic events, such as sudden increases, decreases, or transmission interruptions, and label the time and magnitude of the events. Statistically analyze the traffic characteristics, such as average traffic, peak traffic, traffic variance, and burstiness.

[0291] Collect resource usage data. Record the specific resources used by the data stream, including the OCS optical path identifier, wavelength number, and time period used, or the EPS path used, the nodes and links traversed, and the bandwidth used. Statistically track dynamic changes in resource usage, such as increases or decreases in resource usage over time, resource switching events, resource contention, and waiting status.

[0292] Collect performance metrics data. Measure the transmission performance of the data stream in real time, including current throughput, end-to-end latency, latency jitter, packet loss rate, and retransmission rate. Record the time series of performance metrics and analyze performance trends and fluctuations. Identify abnormal performance events, such as sudden increases in latency, sudden increases in packet loss rate, and sudden decreases in throughput. These anomalies may reflect changes in network conditions or interactions between data streams.

[0293] Collect correlation data from multiple data streams. This includes collecting data not only from individual data streams but also from the relationships between multiple data streams. Record sets of data streams transmitted simultaneously in the network and analyze their temporal overlap. Record data streams sharing the same resources, such as those using the same link, passing through the same node, or competing for the same wavelength. Record business relationships between data streams, such as multiple sub-streams belonging to the same application task, or preceding and subsequent streams with dependencies.

[0294] Construct a behavioral dataset. Integrate all collected data to form a structured behavioral dataset. The dataset is organized in time series format, recording the status and behavioral information of all data streams at each time point. The dataset contains multiple dimensions, each corresponding to a monitoring indicator or feature, forming high-dimensional time series data.

[0295] Step G: Based on the traffic change, resource usage, and performance index data, time series analysis is applied to extract behavioral patterns, and Granger causality test is used to identify the inter-flow influence relationships in the behavioral patterns to obtain a preliminary causal association network.

[0296] In this embodiment, the causal relationship identification module applies time series analysis and causal inference methods to extract the causal influence relationship between data streams from behavioral data and construct a causal association network.

[0297] Time series analysis methods are applied to extract behavioral patterns. Pattern recognition and feature extraction are performed on the traffic time series of each data stream. Autocorrelation analysis is used to identify the periodicity and autocorrelation of the traffic; for example, a data stream may show a peak every 5 seconds, exhibiting a periodic fluctuation pattern. Spectral analysis is used to identify the main frequency components of the traffic, distinguishing between high-frequency fluctuations and low-frequency trends. A change point detection algorithm is used to identify abrupt changes in the traffic time series, marking the moments when significant changes in the traffic pattern occur.

[0298] Extract the coordination patterns between multiple data streams. Analyze the traffic time series of multiple data streams to identify their correlations and synchronicities. Calculate the cross-correlation function between the data streams to identify data stream pairs with time lag relationships. For example, if the traffic change of data stream A leads the traffic change of data stream B by approximately 2 seconds, a causal relationship may exist where A influences B.

[0299] Granger causality tests are used to identify inter-stream influence relationships. A Granger causality test is a statistical method used to determine whether one time series has predictive power for another, thus inferring a causal relationship. For data streams A and B, two time series prediction models are constructed: Model 1 uses only historical data of B to predict future values ​​of B; Model 2 uses historical data of both B and A to predict future values ​​of B. If the prediction accuracy of Model 2 is significantly higher than that of Model 1, it indicates that historical information of A helps predict the future of B, and therefore a Granger causal relationship exists between A and B.

[0300] Perform Granger causality tests on all data stream pairs. Calculate the Granger causality test statistic and significance level for each data stream pair. Set a significance threshold, for example, a p-value less than 0.05. When the test result is significant, a causal relationship is considered to exist; when the test result is not significant, a causal relationship is considered to be absent or weak.

[0301] Construct a preliminary causal relationship network. Using data flows as nodes and causal relationships as directed edges, construct a directed graph. If a Granger causal relationship exists between data flow A and data flow B, add a directed edge from A to B in the graph. This preliminary causal relationship network reflects the topology of influence relationships between data flows, providing a structural foundation for subsequent in-depth analysis.

[0302] Addressing spurious causal relationships. Granger causality tests may identify spurious causal relationships, such as two data streams both being influenced by a third common factor, causing them to appear correlated, but in reality, there is no direct causal relationship. Using conditional Granger causality tests or partial correlation analysis, controlling for the influence of other variables, can identify genuine direct causal relationships, remove spurious relationships, and improve the accuracy of causal networks.

[0303] In one embodiment, time series analysis is a time-dimensional data analysis technique used to extract time-dependent regular patterns from the historical transmission behavior of graph data streams. In multi-stream collaboration scenarios, this method identifies behavioral characteristics such as periodic fluctuations, trend changes, and sudden events by analyzing time series data of traffic changes, resource usage, and performance indicators. Granger causality testing is a statistical hypothesis testing method used to determine whether one time series has a predictive effect on another time series, i.e., whether historical information of one variable helps predict the future value of another variable. In graph data stream analysis, Granger causality testing is used to identify the influence relationships between different data streams, determining whether changes in the behavior of stream A will lead to subsequent changes in stream B, thereby constructing a network structure reflecting causal relationships between streams.

[0304] The core principle of time series analysis is to model the transmission behavior of data streams as a stochastic process that evolves over time, revealing the intrinsic structure and dynamic characteristics of the data through statistical methods. For a single data stream, time series analysis first performs a stationarity test to determine whether the statistical properties of the sequence remain stable over time. The mean, variance, and autocovariance of a stationary sequence do not change over time, making it easier to establish a stable predictive model. Non-stationary sequences can be transformed into stationary sequences using methods such as differencing, logarithmic transformation, or trend decomposition.

[0305] Behavioral pattern extraction employs techniques such as autoregressive moving average models, seasonal decomposition models, and spectral analysis. The autoregressive moving average model represents the current observation as a linear combination of observations from several past times plus a random error term; model parameters are determined through maximum likelihood estimation or least squares. Seasonal decomposition breaks down the time series into trend, seasonal, and random components. The trend component reflects the long-term direction of change, the seasonal component reflects periodic fluctuations, and the random component reflects unpredictable noise. Spectral analysis transforms the time-domain signal to the frequency domain using Fourier transform, identifies the dominant frequency components, and reveals the periodic characteristics of the data.

[0306] The theoretical basis of Granger causality testing is conditional probability and information gain. For two time series X and Y, if historical information about X significantly improves the prediction accuracy of future values ​​of Y, then X is said to be a Granger cause of Y. The testing process constructs two autoregressive models: a restricted model uses only historical values ​​of Y to predict future values ​​of Y, and an unrestricted model uses historical values ​​of both Y and X to predict future values ​​of Y. By comparing the prediction errors of the two models, the F-statistic or likelihood ratio statistic is calculated to determine whether introducing historical information about X significantly reduces the prediction error. If the statistic exceeds a critical value, the null hypothesis is rejected, and X is considered to have a Granger causal relationship with Y.

[0307] Performing Granger causality tests requires determining the lag order, i.e., how many historical moments to consider. A lag order that is too small will miss important time dependencies, while a lag order that is too large will introduce too many parameters, leading to overfitting and increased computational complexity. The choice of lag order is usually based on information criteria, such as the Akaike information criterion or the Bayesian information criterion, selecting the order that minimizes the criterion value. Information criteria balance goodness of fit and complexity, penalizing models with too many parameters.

[0308] The preliminary causal network is constructed based on the results of pairwise Granger causality tests. For a system containing N data flows, N*(N-1) tests are required to cover all possible directed flow pairs. The test results are represented by an adjacency matrix, where an element A equal to 1 indicates that a Granger causal relationship exists between flow i and flow j, and an element A equal to 0 indicates that no such relationship exists. The adjacency matrix defines the topology of the directed graph, where nodes represent data flows and directed edges represent causal relationships. The preliminary causal network may contain spurious causal relationships, i.e., statistical significance due to common driving factors or indirect influences, which require correction through subsequent conditional independence tests and network refinement steps.

[0309] The implementation of time series analysis and Granger causality test consists of five stages: data acquisition and preprocessing, stationarity test and transformation, behavioral pattern extraction, Granger causality test and causal network construction.

[0310] During the data acquisition and preprocessing phase, the system obtains transmission behavior data from multiple graph data streams from the network monitoring system. The time granularity of data acquisition is determined according to application requirements, typically at the second, minute, or hour level. For each data stream, the acquired metrics include instantaneous traffic, cumulative traffic, resource utilization, transmission latency, packet loss rate, and throughput. Data preprocessing includes missing value imputation, outlier detection, and data smoothing. Missing value imputation uses linear interpolation, spline interpolation, or model-based interpolation methods to estimate missing values ​​based on observations at adjacent times. Outlier detection uses statistical methods such as the three-standard-deviation method or density-based methods such as the local anomaly factor algorithm to identify data points deviating from the normal range. Outlier handling strategies include deletion, replacement, or retention and labeling, determined based on the cause and impact of the anomaly. Data smoothing uses moving average or exponential smoothing methods to reduce the impact of high-frequency noise and highlight the main trends in the data.

[0311] In the stationarity testing and transformation phase, the system performs unit root tests on each time series. Commonly used methods include the augmented Dickey-Fuller test and the Phillips-Peron test. The null hypothesis of the augmented Dickey-Fuller test is that the series has a unit root, i.e., it is non-stationary, and the alternative hypothesis is that the series is stationary. The test statistic is calculated based on the first-difference regression model of the series. If the statistic is less than the critical value, the null hypothesis is rejected, and the series is considered stationary. For non-stationary series, differencing methods are used to eliminate trends and seasonality. First-order differencing calculates the difference between observations at adjacent times, and second-order differencing differs the first-order difference result again. Seasonal differencing calculates the difference between the current time and a time a seasonal cycle ago; for example, for daily periodic data, the lag of seasonal differencing is 24 hours. The differrated series usually meets the stationarity requirement. If it is still non-stationary, logarithmic transformation or Box-Cox transformation can be applied to further stabilize the variance.

[0312] In the behavior pattern extraction phase, the system establishes an autoregressive moving average model for the stationary time series. The model order is determined by the autocorrelation function (AFC) and partial autocorrelation function (PIF) plots. The AFC measures the correlation between the series and its lagged versions, while the PIF measures the correlation after removing the influence of intermediate time points. An exponentially decaying AFC and a truncated PIF after lag p indicate suitability for the autoregressive model AR(p). A truncated AFC and exponentially decaying PIF after lag q indicate suitability for the moving average model MA(q). Both AFC and PIF exhibit exponential decay, indicating suitability for the autoregressive moving average model ARMA(p,q). Model parameter estimation employs the maximum likelihood method or conditional least squares method, using iterative optimization algorithms to find parameter values ​​that maximize the likelihood function or minimize the sum of squared residuals.

[0313] Seasonal decomposition employs the STL decomposition method, which is a seasonal trend decomposition based on locally weighted regression. This method estimates the trend component and seasonal component separately through an iterative process. First, a preliminary trend is extracted using a moving average. Then, the trend is subtracted from the original sequence to obtain a detrended sequence. The detrended sequence is grouped according to the seasonal cycle to calculate the seasonal factor. The seasonal component is then subtracted from the detrended sequence to obtain the residual. The above steps are repeated until convergence, yielding the final trend, seasonal, and residual components. The periodicity of the seasonal component is used to predict future periodic fluctuations, while the trend component is used to predict the long-term direction of change.

[0314] Spectral analysis converts a time-domain sequence into a frequency-domain representation using a Fast Fourier Transform (FFT). The spectrogram displays the power density of different frequency components, with the frequency corresponding to the peak indicating the dominant period of the sequence. For example, if the spectrogram shows a peak at 0.04 Hz, corresponding to a period of 25 seconds, it indicates that the data stream exhibits a regular fluctuation every 25 seconds. After identifying the dominant period, a periodic prediction model can be constructed, or the periodic effect can be considered in subsequent analyses.

[0315] Furthermore, critical impact path identification is an algorithm based on graph theory and network analysis used to discover causal propagation sequences that significantly influence system behavior in weighted temporal causal graphs. A critical impact path is a directed path from a source node to a target node, where the cumulative causal impact is strongest and contributes the most to changes in the target node's behavior. A propagation chain is a chain-like structure formed by multiple consecutive causal relationships, reflecting the cascading transmission of influence between multiple data streams. In multi-stream collaborative scenarios of graph data streams, identifying critical impact paths and propagation chains helps understand the dynamic behavior of the system, predict the chain reactions of decision adjustments, and optimize resource allocation strategies. The theoretical basis of critical impact path identification is graph path analysis and centrality measurement. In a weighted temporal causal graph, nodes represent data streams, directed edges represent Granger causal relationships, and edge weights reflect the strength and propagation delay of causal influence. Influence strength is usually quantified using regression coefficients or correlation coefficients, representing the average change in the target stream when the source stream changes by one unit. Propagation delay represents the time required for the influence to travel from the source stream to the target stream, corresponding to the most significant lag order in Granger causality tests.

[0316] Path influence assessment comprehensively considers path length, edge weights, and topology. Path length refers to the number of edges contained in the path, reflecting the number of steps required for influence propagation. Shorter paths generally have a more direct impact, while longer paths may attenuate due to multiple propagations. There are several ways to accumulate edge weights, including multiplicative accumulation and additive accumulation. Multiplicative accumulation assumes that the influence gradually decays along the path, and the total influence of the path equals the product of the weights of all edges on the path. Additive accumulation assumes that the influence accumulates along the path, and the total influence of the path equals the sum of the weights of all edges on the path. In practical applications, the appropriate accumulation method is selected based on the influence propagation mechanism.

[0317] Centrality measures are used to identify critical nodes and edges in a network. Betweenness centrality measures the importance of a node or edge as a mediator in the network, defined as the proportion of shortest paths through that node or edge to the total number of shortest paths. Nodes with high betweenness centrality are bottlenecks or hubs in the network; their failure significantly impacts network connectivity and information propagation. Proximity centrality measures the average distance from a node to all other nodes in the network; the shorter the distance, the higher the centrality, indicating that the node can quickly influence or be influenced. Eigenvector centrality considers the importance of a node's neighbors; a node has high centrality if it is connected to multiple important nodes.

[0318] The identification of propagation chains is based on path enumeration and chain structure analysis. A propagation chain is a special type of path where each node has an out-degree of 1, forming a linear causal sequence. The starting point of the chain is the source of influence, and the ending point is the ultimately affected target. The length of the propagation chain reflects the depth of the influence's propagation, and the order of the nodes on the chain reflects the propagation sequence of the influence. Identifying propagation chains helps to trace the source and propagation process of influence, and predict the cascading effects of a change in the behavior of a node on downstream nodes.

[0319] Algorithms for identifying critical impact paths include shortest path algorithms, maximum weight path algorithms, and k-optimal path algorithms. Shortest path algorithms, such as Dijkstra's algorithm and the Bellman-Ford algorithm, find the path from the source node to the target node with the fewest edges or the smallest sum of weights. Maximum weight path algorithms find the path with the largest product of weights or the largest sum of weights; this is usually achieved by modifying the shortest path algorithm by negating or taking the logarithm of the edge weights, transforming it into a shortest path problem. K-optimal path algorithms, such as Yen's algorithm, find the top k paths by weight, providing multiple candidate paths for decision-makers to choose from.

[0320] The identification of key impact paths and propagation chains is divided into five stages: weight calculation, path enumeration, path evaluation, key path screening, and propagation chain extraction.

[0321] During the weight calculation phase, the system assigns a weight to each edge of the temporal causal graph. The weight includes two dimensions: influence strength and propagation delay. The influence strength is calculated based on the regression coefficients of the Granger causality test. For edge (i, j), the influence strength w is equal to the sum of the absolute values ​​of the lag coefficients of flow i in the unconstrained model.

[0322] The propagation delay is determined based on the most significant lag order. For each lag order τ, the t-statistic of the corresponding regression coefficient βτ is calculated, and the lag order with the largest absolute value of the t-statistic is selected as the propagation delay d. The propagation delay reflects the typical time it takes for the influence to be transmitted from flow i to flow j. If the t-statistics of multiple lag orders are similar, a weighted average delay can be calculated, with the weights being the absolute values ​​of the regression coefficients of each lag order.

[0323] During the path enumeration phase, the system uses either depth-first search (DFS) or breadth-first search (BFS) algorithms to traverse the causal graph, enumerating all possible paths from a given source node to the target node. DFS starts from the source node and delves as deep as possible along a path until the target node is reached or further exploration is impossible, then backtracks to explore other paths. BFS traverses level by level, first exploring all nodes one step away from the source node, then nodes two steps away, and so on. To avoid path explosion, a path length cap is set, for example, a maximum of 5 edges; paths exceeding this cap are not considered.

[0324] The path enumeration process requires handling loop problems. If the causal graph contains loops—that is, paths starting from node i, passing through several nodes, and returning to i—then infinitely long paths may exist. Loop detection is achieved by maintaining a set of visited nodes; if a node is already included in the current path, that node is not visited again to avoid duplication. The existence of loops indicates a feedback mechanism between data flows: flow i influences flow j, and flow j, in turn, influences flow i, forming a mutually influential cycle. Feedback loops are significant in system dynamics, potentially leading to oscillations or stability, and require separate analysis.

[0325] During the path evaluation phase, the system calculates the overall influence of each enumerated path. The calculation method for path influence is selected based on the application scenario. For scenarios with diminishing influence, multiplicative accumulation is used: the influence I(P) of path P is equal to the product of the weights of all edges on the path, i.e., I(P) = W1 * W2 * ... * Wn, where Wi is the overall weight of the i-th edge on the path. Multiplicative accumulation reflects the gradual decay of influence during propagation; for each edge encountered, the influence is multiplied by the weight of that edge.

[0326] For scenarios involving cumulative effects, additive accumulation is used: I(P) = W1 + W2 + ... + Wn. Additive accumulation is suitable for situations where effects can be superimposed, such as when changes in the flow of multiple flows collectively affect the target flow, and the total effect is the sum of the individual effects. Additive accumulation can also be weighted, assigning different weights based on the position of edges in the path; for example, edges at the beginning of the path have higher weights, reflecting the importance of earlier effects.

[0327] The propagation delay of a path is equal to the sum of the propagation delays of all edges along the path, i.e., T(P) = d1 + d2 + ... + dn. The propagation delay reflects the total time required for an influence to travel from the source node to the target node, and is an important parameter for predicting and adjusting the timescale of chain reactions.

[0328] During the critical path selection phase, the system ranks all paths based on their impact and propagation delay, selecting the top-ranked paths as critical impact paths. Ranking can be based on a single indicator or a combination of multiple indicators. Based on impact ranking, the k paths with the highest I(P) are selected, as these paths have the strongest impact on the target node. Based on delay ranking, the k paths with the lowest T(P) are selected, as these paths propagate their impact the fastest. The multi-indicator approach uses a weighted scoring system, defining the path score S(P) as equal to α*I(P) - β*T(P), where α and β are weighting coefficients determined based on the importance placed on impact and delay. The k paths with the highest S(P) are selected as critical paths.

[0329] The number of critical paths, k, is determined based on application requirements and network size. For small networks, k can be set to 3 to 5, providing a small number of the most important paths for analysis. For large networks, k can be set to 10 to 20, covering more influence patterns. The selection of critical paths can also consider path diversity, avoiding the selection of highly similar paths. Path similarity is measured by path overlap, which is the proportion of edges shared by two paths to the total number of edges in the path. During the screening process, if the overlap between a new candidate path and a selected path exceeds a threshold, such as 50%, that path is skipped, and the next candidate path is selected, ensuring that the selected critical paths cover different influence mechanisms.

[0330] In the propagation chain extraction phase, the system identifies linear propagation structures in the causal graph. Propagation chain identification is based on topological analysis, searching for paths that satisfy the following conditions: except for the starting and ending points, each intermediate node on the path has an in-degree and out-degree of 1, meaning each node has only one predecessor and one successor. This structure indicates that influences propagate along a unique path, without branches or convergences.

[0331] The algorithm for extracting propagation chains starts from each node and proceeds along the outgoing edges. If the outgoing degree of the current node is 1, it continues along the unique outgoing edge until it reaches a node with an outgoing degree other than 1 or reaches the endpoint. The sequence of nodes traversed is recorded, forming a propagation chain. This process is repeated for all nodes to obtain all propagation chains in the network. Propagation chains may overlap, meaning multiple chains may share some nodes; the system needs to deduplicate or merge overlapping chains.

[0332] The importance assessment of a propagation chain is based on the chain's length, total influence, and the importance of the nodes it covers. Longer chains reflect deeper propagation of influence, potentially leading to long-distance cascading effects. Total influence equals the product or sum of the weights of all edges in the chain, reflecting the overall strength of the chain's impact. The importance of the nodes covered is measured by their centrality or business criticality; a chain is more important if it contains critical business flows.

[0333] The identified key impact paths and propagation chains are used to construct a causal propagation model. This model is a simplification and abstraction of the original causal graph, retaining the most important causal relationships and propagation structure while removing secondary edges and nodes. Using the key paths and propagation chains as a framework, the model labels the impact, propagation delay, and data flow involved in each path, providing a structured knowledge representation for subsequent chain reaction prediction and collaborative decision-making.

[0334] Step H: Calculate the propagation delay and propagation strength using the preliminary causal relationship network, and output a weighted temporal causal graph.

[0335] In this embodiment, the causal relationship quantification module performs in-depth analysis on the preliminary causal relationship network, quantifies the temporal and intensity characteristics of causal relationships, and constructs a weighted time-series causal graph.

[0336] Calculate the propagation delay. For each directed edge in the causal network, i.e., each pair of data streams with a causal relationship, calculate the time required for the impact to propagate from the source data stream to the target data stream. The delay can be obtained from the peak position of the cross-correlation function; the time lag corresponding to the peak is the propagation delay. For example, if the cross-correlation function of data streams A and B reaches its peak at a lag of 2 seconds, it means that a change in A requires a 2-second propagation delay to affect B.

[0337] Propagation delay reflects the time scale of causal effects and is of great significance for prediction and decision-making. Short-delay causal relationships indicate rapid propagation of effects, requiring timely responses; long-delay causal relationships indicate slow propagation of effects, with a longer reaction time window.

[0338] Calculate the intensity of propagation. This involves quantifying the strength of the causal relationship, i.e., the extent of the influence of the source data stream on the target data stream. Propagation intensity can be measured from multiple perspectives. One method is to calculate the magnitude of the Granger causality test statistic; a larger statistic indicates a stronger causal relationship. Another method is to calculate the peak value of the cross-correlation coefficient; a higher peak value indicates a stronger correlation between the two data streams and a greater causal influence. Information theory methods can also be used to calculate the amount of information transferred from the source data stream to the target data stream, such as mutual information or transit entropy; a larger amount of information transferred indicates a stronger causal influence.

[0339] Normalized propagation strength maps strength values ​​to a range of 0 to 1 for easier comparison and visualization. A strength of 1 indicates the strongest causal influence, while a strength close to 0 indicates a weak or non-existent causal influence.

[0340] Construct a weighted temporal causal graph. Based on the preliminary causal relationship network, add two weight attributes to each directed edge: propagation delay and propagation strength. The directed edge is labeled as: from node A to node B, with a delay of T seconds and a strength of S. The weighted temporal causal graph not only represents the causal relationship topology between data flows but also quantifies the temporal and strength characteristics of causal relationships, providing richer information.

[0341] Visualize the time-series cause-effect graph. Using graph visualization tools, the time-series cause-effect graph is displayed in an intuitive graphical form. Nodes represent data flows, and node size reflects the volume or importance of the data flow. Directed edges represent causal relationships; edge thickness reflects propagation strength, with thick edges indicating strong causal relationships and thin edges indicating weak causal relationships. Edge color or labels reflect propagation latency; for example, different colors represent short, medium, and long latency. The visualization helps operations and maintenance personnel intuitively understand the complex causal relationships between data flows, assisting in decision-making.

[0342] Step 1: Identify the key influence paths and propagation chains in the weighted temporal causal graph to obtain the causal propagation model.

[0343] In this embodiment, the critical path identification module analyzes the weighted temporal causal graph, identifies the key influence paths and propagation chains, constructs a causal propagation model, and reveals the propagation mechanism of influence in the data flow network.

[0344] Identify critical impact paths. In a time-series cause-effect graph, influences may propagate along causal chains, from one data stream to another, and then to a third, forming impact paths. Critical impact paths are those that have the greatest impact on the overall system performance. Various methods can be used to identify critical paths.

[0345] One approach is based on path strength. The total strength of a path is calculated as the product or weighted sum of the propagation strengths of all edges along that path. The path with the highest strength is the critical path. For example, on the path A to B to C, the strength from A to B is 0.8, and the strength from B to C is 0.7, so the total strength is 0.8 multiplied by 0.7, which equals 0.56. The total strengths of all paths are compared, and the paths with the highest total strengths are selected as the critical paths.

[0346] Another approach is based on path coverage. This involves identifying the paths that traverse the most nodes or affect the most data flows. These paths have a wide reach within the network and significantly impact the overall system state.

[0347] It can also be based on the latency characteristics of the paths. Identifying the shortest end-to-end latency paths, where effects propagate quickly, is crucial for real-time decision-making and rapid response. Alternatively, identifying the longest end-to-end latency paths, where effects propagate slowly, may lead to delay effects and cumulative impacts.

[0348] Identify propagation chains. A propagation chain refers to a series of data streams connected in a causal sequence, forming a transmission sequence of influence. Analyze the topology of the propagation chain to identify different patterns such as chain structures, tree structures, and network structures. A chain structure indicates that the influence propagates linearly along a single path; a tree structure indicates that one source data stream influences multiple target data streams, with the influence spreading in a fan-shaped pattern; a network structure indicates that multiple data streams influence each other, forming complex feedback and coupling relationships.

[0349] Identify key nodes. In a time-series causal graph, certain nodes possess special importance. Central nodes are those with high degree or in-degree, meaning they influence or are influenced by many other data flows. These nodes occupy a central position in the causal network, and intervention in them has a wide-ranging impact. Bottleneck nodes are those located on the critical path, whose performance or state significantly affects path propagation. Anomalies in bottleneck nodes can block or weaken the propagation of influence. Bridging nodes are nodes connecting different sub-networks or communities, acting as bridges in network connectivity and influence propagation.

[0350] Construct a causal propagation model. A causal propagation model integrates information such as key influence paths, propagation chains, and key nodes to form a systematic understanding of the causal relationship network of data flows. The model can be represented mathematically, for example, using state transition equations to describe how influence propagates from one data flow to another, and using a propagation matrix to represent the influence relationships between all nodes in the network. The model can also be represented graphically or textually, clearly describing the main causal relationships and propagation mechanisms.

[0351] Causal propagation models reveal the implicit relationships and dynamic behaviors in data flow networks, providing a theoretical basis for predicting system responses and optimizing decision-making strategies.

[0352] Step J: Based on the causal propagation model, predict the chain reaction of decision adjustment to achieve optimal decision-making through multi-stream collaboration.

[0353] In this embodiment, the decision optimization module uses a causal propagation model to predict the chain effect of exchange decision adjustments on multiple data streams, formulates a collaborative decision-making strategy that takes causal relationships into account, and optimizes the overall performance of the system.

[0354] The direct impact of predictive decision adjustments. When adjusting the switching decision for a data stream, such as switching data stream A from the EPS path to the OCS path, we first analyze the direct impact of this adjustment on data stream A itself. After switching to OCS, data stream A may gain higher bandwidth and lower latency, improving transmission performance; however, the establishment of the OCS optical path requires time overhead, which may lead to temporary transmission interruptions or delays.

[0355] Predicting the indirect impact of decision adjustments. According to the causal propagation model, changes in data flow A may affect other data flows. If the causal graph shows a causal relationship between data flow A and data flow B, with a propagation delay of 2 seconds and a propagation strength of 0.7, then it can be predicted that 2 seconds after the decision adjustment, the state of data flow B will be affected, with the degree of impact proportional to the propagation strength of 0.7. Specifically, the impact may manifest as follows: if the traffic of data flow A increases, the traffic of data flow B may also increase; if data flow A occupies a shared resource, data flow B may face resource contention and performance degradation.

[0356] Predicting chain reactions. The impact not only propagates at the first level but may also continue to spread along causal chains. Data flow A affects data flow B, which in turn affects data flow C, forming a chain reaction. Based on the causal propagation model, calculate the cumulative effect of the impact propagating along the path. For example, on the path A to B to C, if the delay from A to B is 2 seconds with an intensity of 0.7, and the delay from B to C is 3 seconds with an intensity of 0.6, then the total delay from A to C is 2 plus 3 equals 5 seconds, and the total intensity is 0.7 multiplied by 0.6 equals 0.42. Predict that 5 seconds after the decision adjustment, data flow C will be affected by an intensity of 0.42.

[0357] Evaluate the overall effect of the decision adjustment. Consider the impact of the decision adjustment on all relevant data streams and calculate the change in overall system performance. Overall performance can be represented by a weighted sum, for example, summing the performance metrics of each data stream after multiplying them by their importance weight. If the decision adjustment improves the performance of high-priority data streams but degrades the performance of low-priority data streams, it is necessary to weigh the pros and cons and determine whether the overall effect is positive or negative.

[0358] Develop collaborative decision-making strategies that consider causal relationships. Traditional decision-making methods often consider individual data streams in isolation, ignoring the causal connections between them. Decision-making methods based on causal propagation models, when adjusting a data stream, simultaneously consider the impact on other data streams, thus formulating collaborative decision-making strategies.

[0359] For example, when allocating OCS resources to a high-priority data flow A, if the causal model shows that data flow A has a strong causal impact on data flows B and C, then while allocating resources to A, resources should also be reserved or adjusted for B and C to prevent changes in A from causing performance degradation in B and C. Alternatively, if it is found that data flow D has a negative impact on multiple other data flows—for example, D consuming a large amount of resources and causing congestion in other flows—rate limiting or path isolation measures can be taken for D to reduce its impact on other flows.

[0360] Optimize decision-making timing. Based on the time delay information of causal propagation, select the optimal time to adjust decisions. If the impact of a decision takes a long time to propagate to critical data streams, adjustments can be made in advance to ensure the impact arrives at the necessary moment. If the impact of a decision propagates quickly, adjustments should be made closer to the required moment to avoid resource waste or state mismatch caused by premature adjustments.

[0361] Achieving optimal decision-making through multi-stream collaboration. By comprehensively considering causal relationships, impact propagation, and chain reactions, the decision-making system elevates from local optimization to global optimization, and from static decision-making to dynamic collaborative decision-making. The system can anticipate the long-term impact of decisions, avoid systemic problems caused by short-sighted decisions, and achieve coordinated operation and optimal overall performance of multiple data streams within a causal network.

[0362] Example 9

[0363] This embodiment adds a new step to Embodiment 1. The method further includes: acquiring historical decision records of the graph data stream, extracting decision scenarios and effect data, constructing a decision tree model to characterize decision rules based on the decision scenarios and effect data, and obtaining a historical decision model; generating multiple sets of hypothetical scenarios based on the historical decision model using the Monte Carlo method to generate a counterfactual scenario set; performing simulation experiments to evaluate the effects of different decisions based on the counterfactual scenario set, and outputting a comparison result of decision effects; calculating the optimal possible performance of each decision point in the comparison result of decision effects, and updating the decision rules and evaluation criteria based on the optimal possible performance to achieve continuous optimization of the decision mechanism.

[0364] Step K: Obtain the historical decision records of the graph data stream, extract the decision scenarios and effect data, and construct a decision tree model based on the decision scenarios and effect data to characterize the decision rules, thereby obtaining the historical decision model.

[0365] In this embodiment, the historical decision analysis module obtains historical decision records from the system log, analyzes past decision processes and results, and constructs a historical decision model to provide a foundation for counterfactual analysis and decision optimization.

[0366] Retrieve historical decision records. Extract all decision records from the decision system's log database over a past period. Each decision record includes: the decision time (the point in time the decision was made); the decision scenario (the system state and environmental information at the time of the decision, including data flow characteristics, network resource status, and service priorities); the decision content (the actual actions taken, such as choosing between the OCS and EPS paths for a data flow, and allocating specific resources); and the decision result (the actual effect after the decision was executed, including transmission performance metrics, resource utilization, and service satisfaction).

[0367] Extract decision-making scenario features. Structure and characterize various information from the decision-making scenario. Decision-making scenario features include: data flow features, such as traffic volume, duration, graph traversal depth, classification type, and confidence level; network status features, such as the number of available wavelengths in the OCS, EPS link utilization, the number of currently transmitted data streams, and network congestion levels; time features, such as the hour, week, and whether it is a peak business period; and historical features, such as recent decision-making history and performance trends.

[0368] Extract decision-making effectiveness data. Quantify the actual effects of decisions and construct effectiveness evaluation indicators. Effectiveness indicators include: performance indicators, such as actual values ​​and deviations from target values ​​for throughput, latency, and packet loss rate; resource efficiency indicators, such as resource utilization, configuration overhead, and resource waste; and business indicators, such as service quality satisfaction rate, user satisfaction, and degree of business value realization. Combine multiple effectiveness indicators into an overall score, for example, using a weighted sum or comprehensive evaluation function, to obtain the effectiveness score for each decision.

[0369] Build a decision tree model to characterize decision rules. A decision tree is an interpretable machine learning model that can clearly represent decision rules in a tree structure. Use historical decision records as training data, decision scenario features as input, and decision content as output labels to train the decision tree model.

[0370] The process of constructing a decision tree is as follows: Starting from the root node, the optimal feature is selected for splitting. The optimal feature is the feature that maximizes the purity of the child nodes. Purity can be measured by indicators such as information gain and Gini coefficient. For example, if splitting according to the feature "whether the traffic size is greater than 100MB" can well divide the data into two categories, one mainly selecting OCS and the other mainly selecting EPS, then this feature is a good splitting feature.

[0371] The process recursively splits each child node, selecting the suboptimal feature to continue splitting, until a stopping condition is met, such as sufficiently high node purity, insufficient node sample size, or the tree depth reaching a limit. Leaf nodes represent the final decision result, such as "choose the OCS path" or "choose the EPS path".

[0372] The trained decision tree model represents the rule patterns of historical decisions in a tree structure. A path from the root node to a leaf node represents a decision rule, such as: "If the traffic size is greater than 100MB, the duration is greater than 10 seconds, and the OCS has available resources, then choose the OCS path." The decision tree model clearly depicts the logic of historical decisions and reveals the behavioral patterns of the decision system in different scenarios.

[0373] The historical decision-making model is obtained. This model includes not only the decision tree structure but also statistical information on decision outcomes. For each leaf node of the decision tree, the effect scores of all historical records of making that decision are calculated, along with the average effect, best effect, worst effect, and effect variance. This statistical information reflects the historical performance of the decision rule and provides a basis for evaluating decision quality.

[0374] Step L: Based on the historical decision-making model, generate multiple sets of hypothetical scenarios using the Monte Carlo method to generate a set of counterfactual scenarios.

[0375] In this embodiment, the counterfactual scenario generation module generates multiple sets of hypothetical decision scenarios based on the historical decision-making model and through the Monte Carlo simulation method, constructing a counterfactual scenario set to explore the possible outcomes of different decision choices.

[0376] Understand the concept of counterfactual analysis. Counterfactual analysis refers to considering historical decisions and hypothesizing "what would have happened if a different decision had been made?" For example, if a data stream was historically assigned to the EPS path, counterfactual analysis might hypothesize "how would performance have changed if that data stream had been assigned to the OCS path?" By comparing the effects of the actual and counterfactual decisions, the merits of the decisions are evaluated, and areas for improvement are identified.

[0377] Identify key decision points. From historical decision records, identify decision points that significantly impact system performance. Key decision points may include: decisions with low performance scores, indicating that the decision may not be optimal and has potential for improvement; scenarios at the decision boundary, where the scenario characteristics are close to the classification boundary and the decision choice is uncertain; and decisions involving high-priority or high-impact data flows, which have a significant impact on overall performance.

[0378] Hypothetical scenarios are generated using the Monte Carlo method. The Monte Carlo method is a simulation technique based on random sampling that explores the possible behaviors of a system through a large number of random trials. For each key decision point, multiple sets of hypothetical scenarios are generated, each set corresponding to one possible decision choice.

[0379] The specific generation process is as follows: The basic characteristics of the decision-making scenario remain unchanged, i.e., data flow characteristics, network status, etc., remain consistent with historical records. The decision content is changed, i.e., different decision actions are assumed. For example, historically, the EPS path was chosen; now, scenario 1 might choose the OCS path, scenario 2 might choose a hybrid path, and scenario 3 might choose delayed transmission, etc.

[0380] For each hypothetical decision, its possible execution process and results need to be simulated. Since the hypothetical decision is not actually executed, its results cannot be directly observed and need to be predicted through a model. Statistical information on effects from historical decision models can be used; for example, assuming the OCS path is chosen, the possible effects of this hypothetical decision can be predicted based on historical statistics of the effects of choosing OCS in similar scenarios.

[0381] Randomness and uncertainty are introduced. The behavior of real-world systems is influenced by various random factors, such as random fluctuations in network conditions and random changes in data flow characteristics. When generating hypothetical scenarios, these random factors are introduced through Monte Carlo sampling. For example, assuming the OCS optical path establishment time follows a certain probability distribution, a specific establishment time value is obtained through random sampling; assuming random jitter in network latency, a specific latency value is obtained through sampling. Through multiple random samplings, multiple sets of hypothetical scenarios are generated, each corresponding to a possible random realization.

[0382] Generate a set of counterfactual scenarios. For each key decision point, generate several sets of counterfactual scenarios, such as 10 or 100 scenarios per decision point. The counterfactual scenario set contains a large number of hypothetical situations, covering possible outcomes under different decision choices and different random conditions, providing rich data for comprehensive decision evaluation.

[0383] Step M: Based on the counterfactual scenario set, perform simulation experiments to evaluate the effects of different decisions and output the comparison results of decision effects.

[0384] In this embodiment, the simulation evaluation module performs simulation experiments based on a counterfactual scenario set, simulates the system behavior and performance under different decision choices, evaluates the effects of various decisions, and outputs comparison results.

[0385] Constructing simulation models. Simulation models are abstractions and simulations of actual systems, capable of reproducing the system's operation in a computer. Simulation models include: a network topology model, representing the network's nodes, links, devices, and other structures; a resource model, representing the resource capacity, configuration characteristics, and performance parameters of the OCS and EPS; a data flow model, representing the data flow's arrival process, transmission characteristics, and service requirements; and a decision model, representing the decision system's classification algorithms, path selection strategies, and resource allocation rules.

[0386] Configure simulation parameters. Based on the counterfactual scenario, configure the parameters of the simulation model. Input the characteristics of the decision scenario into the simulation model, such as the data flow volume and network resource status. Define the hypothetical decision content, such as selecting a specific switching path for a data flow. Set random parameters, such as random values ​​for optical path establishment time and network latency.

[0387] Perform simulation experiments. Run the simulation model to simulate the data flow transmission process. The simulation process includes: data flow arrival, triggering the decision system to classify and select paths; based on the decision results, configuring switching paths, establishing OCS optical paths or setting EPS routes; data flow transmission through the selected path, and the simulation model calculating performance indicators such as transmission time, throughput, and latency; resource occupation and release, updating network resource status; and recording various events and performance data during the simulation process.

[0388] For each counterfactual scenario, a simulation experiment is performed to obtain the performance results under that scenario. Due to the introduction of randomness, the same decision choice may produce different results under different random conditions. Therefore, for each decision choice, multiple simulation experiments are performed, for example, 100 times, to obtain the distribution of performance results.

[0389] Evaluate the effectiveness of different decisions. For each key decision point, compare the simulation results of the actual decision and various counterfactual decisions. Calculate statistical indicators such as average performance, best performance, worst performance, and performance variance for each decision. For example, the actual decision choosing the EPS path has an average latency of 50 milliseconds; the counterfactual decision choosing the OCS path has an average latency of 20 milliseconds, showing a significant performance improvement.

[0390] Output the comparison results of decision effects. The comparison results are presented in the form of tables, charts, or reports, clearly showing the performance differences of different decision choices. For example, create a comparison table listing the key performance indicators of each decision choice; draw a performance distribution chart to show the performance probability distribution of each decision choice; generate an analysis report summarizing which decision choices performed well, which performed poorly, and the reasons for the performance differences.

[0391] Step N: Calculate the optimal possible performance of each decision point in the comparison results of decision effects, and update the decision rules and evaluation criteria based on the optimal possible performance to achieve continuous optimization of the decision-making mechanism.

[0392] In this embodiment, the decision optimization module identifies the optimal decision based on the results of counterfactual analysis, updates the decision rules and evaluation criteria, and achieves continuous improvement of the decision-making mechanism.

[0393] Calculate the optimal possible performance for each decision point. For each critical decision point, identify the best-performing decision choice among all counterfactual simulation results. Optimal performance can be measured by a single metric, such as lowest latency or highest throughput, or by a comprehensive score, such as a weighted multi-objective optimization score. Record the optimal decision choice and its corresponding performance value as the optimal possible performance for that decision point.

[0394] Calculate the decision regret value. The decision regret value is defined as the difference between the performance of the actual decision and the optimal possible performance, reflecting the optimization space of the decision. For example, if the delay of the actual decision is 50 milliseconds and the delay of the optimal possible performance is 20 milliseconds, the regret value is 50 - 20 = 30 milliseconds. The larger the regret value, the further the actual decision is from the optimal decision, and the greater the potential for improvement.

[0395] Identify decision rules that need improvement. Calculate the regret value for all decision points and identify decision scenarios with high regret values; these scenarios are key areas for decision optimization. Analyze the common characteristics of these scenarios to identify the reasons for suboptimal decisions. For example, it was found that decision regret values ​​are generally high in scenarios where traffic volume is close to the threshold boundary, indicating that the current threshold setting may not be reasonable and needs adjustment.

[0396] Update the decision rules. Based on the results of counterfactual analysis, modify the parameters and rules of the decision tree model or decision algorithm. For example, if it is found that in a certain scenario, choosing OCS always performs better than choosing EPS, but the current decision rule favors EPS, then modify the rules to increase the bias towards choosing OCS. Specific modification methods could include: adjusting the splitting threshold of the decision tree, such as increasing the traffic size threshold from 100MB to 120MB; modifying feature weights to increase the importance of certain features and decrease the importance of others; and adding new decision rules to set specific decision strategies for particular scenarios.

[0397] Update evaluation criteria. Counterfactual analysis may reveal shortcomings in current evaluation criteria. For example, current criteria may only focus on delays, but counterfactual analysis may find that some decisions, while having low delays, result in poor resource utilization, leading to overall inefficiency. Therefore, updating evaluation criteria by increasing the weight of resource utilization indicators or introducing new evaluation dimensions, such as energy consumption and cost, can make the evaluation more comprehensive and reasonable.

[0398] Verify the optimization effect. Apply the updated decision rules to historical data or new simulation scenarios to evaluate the performance of the optimized decisions. Compare the performance metrics before and after optimization to verify whether the update of the decision rules has brought about actual improvement. If the optimization effect is significant, formally deploy the updated decision rules; if the effect is not obvious or there are negative impacts, further analyze the reasons and adjust the optimization strategy.

[0399] Establish a continuous optimization mechanism. Counterfactual analysis and decision optimization are not one-off processes, but rather continuous cycles. Perform counterfactual analysis regularly, such as weekly or monthly, to generate new counterfactual scenarios based on the latest historical data and system performance, evaluate the merits of current decisions, and identify new opportunities for improvement. Through continuous counterfactual learning and decision optimization, the system's decision-making capabilities continuously improve, gradually approaching the optimal decision-making strategy, thus achieving continuous optimization of the decision-making mechanism and continuous performance improvement.

[0400] Example 10

[0401] This embodiment provides a hybrid switching intelligent classification system based on graph data stream features, comprising: an acquisition module for acquiring graph data streams, parsing the graph data streams, extracting multi-dimensional feature parameters including traffic size, duration, packet interval, and graph traversal depth, and constructing a graph data stream feature profile model; a classification module for classifying the graph data streams based on the graph data stream feature profile model using a multi-dimensional feature threshold determination algorithm, and outputting the graph data stream classification result; a hybrid switching module for selecting an optical circuit switching (OCS) path or an electrical packet switching (EPS) path as the actual switching path for hybrid switching of the graph data streams based on the graph data stream classification result using a switching path decision algorithm; an optimization module for collecting performance data to evaluate the actual switching effect, and performing machine learning optimization on the classification threshold and decision rules based on the performance data to update the classification parameters and determination rules; and a transmission module for performing dynamic matching of the graph data streams with network resources according to the updated classification parameters and determination rules to achieve intelligent transmission of the graph data streams.

[0402] In this embodiment, the functions of each module of the system correspond to the corresponding steps in the aforementioned method embodiments. The acquisition module implements the function of S1 in embodiment 1, the classification module implements the function of S2, the hybrid exchange module implements the function of S3, the optimization module implements the function of S4, and the transmission module implements the function of S5.

[0403] The acquisition module includes a data capture unit, a protocol parsing unit, a feature extraction unit, and a profile construction unit. The data capture unit captures the raw graph data stream through a network interface; the protocol parsing unit parses the data packets and extracts the stream identification information; the feature extraction unit calculates temporal features, scale features, and graph structure features; and the profile construction unit constructs a graph data stream feature profile model through a feature fusion algorithm.

[0404] The classification module includes a threshold initialization unit, a weight configuration unit, a scoring calculation unit, a type partitioning unit, and a confidence calculation unit. The threshold initialization unit sets the initial thresholds for the multi-dimensional judgment parameters; the weight configuration unit configures feature weights according to the network environment and application scenario; the scoring calculation unit performs weighted scoring on the data stream features; the type partitioning unit classifies the data stream based on the comprehensive judgment score; and the confidence calculation unit adds confidence information to the classification results.

[0405] The hybrid switching module includes a resource monitoring unit, a decision execution unit, an OCS configuration unit, and an EPS configuration unit. The resource monitoring unit monitors the resource status of the OCS and EPS in real time; the decision execution unit selects the switching path based on the classification results and resource status; the OCS configuration unit is responsible for the establishment, management, and teardown of optical paths; and the EPS configuration unit is responsible for optimizing routing policies and configuring QoS parameters.

[0406] The optimization module includes a performance acquisition unit, a model training unit, and a parameter update unit. The performance acquisition unit collects performance data from the data stream transmission; the model training unit builds and trains the machine learning model, outputting an optimized parameter set; and the parameter update unit updates the classification threshold and decision rules based on the optimized parameter set.

[0407] The transmission module includes a policy formulation unit, a configuration optimization unit, and a resource matching unit. The policy formulation unit formulates data flow processing strategies based on updated parameters; the configuration optimization unit performs batch configuration and pre-configuration, generating optimized configuration instructions; and the resource matching unit performs dynamic matching of data flows with network resources to achieve intelligent transmission.

[0408] The system in this embodiment may further include a forward planning module, a causal analysis module, and a counterfactual learning module, corresponding to the extended functions in embodiments 7, 8, and 9, respectively. The forward planning module includes a historical analysis unit, a state prediction unit, a demand calculation unit, a path planning unit, and a collaborative optimization unit; the causal analysis module includes a behavior acquisition unit, a relationship identification unit, a graph construction unit, a path identification unit, and a decision optimization unit; the counterfactual learning module includes a historical modeling unit, a scene generation unit, a simulation evaluation unit, and a rule optimization unit.

[0409] The system in this embodiment achieves intelligent classification and optimized transmission of mixed graph data streams through modular design. The modules work together to form a complete intelligent transmission system.

[0410] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hybrid exchange intelligent classification method based on graph data stream features, characterized in that, include: Obtain the graph data stream, parse the graph data stream, extract multi-dimensional feature parameters including traffic size, duration, data packet interval and graph traversal depth, and construct a graph data stream feature profile model; Based on the graph data stream feature profiling model, a multi-dimensional feature threshold determination algorithm is applied to classify the graph data stream and output the graph data stream classification result; Based on the graph data stream classification results, the optical circuit switching (OCS) path or electrical packet switching (EPS) path is selected as the actual switching path through the switching path decision algorithm to perform the hybrid switching of the graph data stream. Performance data is collected to evaluate the actual exchange effect. Based on the performance data, classification parameters and decision rules are optimized using machine learning to update the classification parameters and decision rules. According to the updated classification parameters and decision rules, dynamic matching of graph data streams and network resources is performed to achieve intelligent transmission of graph data streams. The transmission behavior of the graph data streams is continuously monitored, and data on traffic changes, resource consumption, and performance indicators are collected. Based on the traffic changes, resource consumption, and performance indicator data, time series analysis is applied to extract behavioral patterns, and Granger causality tests are used to identify the inter-flow influence relationships in the behavioral patterns to obtain a preliminary causal association network. The preliminary causal association network is then used... The network computation affects propagation delay and propagation intensity, outputting a weighted temporal causal graph; key influence paths and propagation chains in the weighted temporal causal graph are identified to obtain a causal propagation model; based on the causal propagation model, the chain reaction of decision adjustments is predicted to achieve optimal decision-making through multi-stream collaboration; historical decision records of the graph data stream are obtained, and decision scenarios and effect data are extracted; based on the decision scenarios and effect data, a decision tree model is constructed to characterize decision rules, resulting in a historical decision model; based on the historical decision model, multiple sets of hypothetical scenarios are generated using the Monte Carlo method to generate a counterfactual scenario set; based on the counterfactual scenario set, simulation experiments are performed to evaluate the effects of different decisions, and the comparison results of decision effects are output. For each decision point, the optimal decision choice is identified from the simulation results of all counterfactual scenarios. Based on the optimal decision choice, the decision rules and evaluation criteria are updated to achieve continuous optimization of the decision-making mechanism.

2. The method according to claim 1, characterized in that, The process involves acquiring a graph data stream, parsing the graph data stream, extracting multi-dimensional feature parameters including traffic size, duration, packet interval, and graph traversal depth, and constructing a graph data stream feature profile model. This includes: acquiring the graph data stream; parsing the graph data stream via a network interface to obtain a preliminary graph data stream set; calculating the packet arrival interval distribution, flow duration, and burstiness index based on the preliminary graph data stream set to obtain a traffic temporal feature vector; calculating the number of bytes, packets, and average packet size of the preliminary graph data stream set and normalizing them to obtain a scale feature vector; analyzing the graph traversal depth, number of accessed nodes, and edge traversal frequency based on the preliminary graph data stream set to obtain a graph structure feature vector; and constructing the graph data stream feature profile model based on the traffic temporal feature vector, the scale feature vector, and the graph structure feature vector using a feature fusion algorithm.

3. The method according to claim 1, characterized in that, Based on the graph data stream feature profiling model, a multi-dimensional feature threshold determination algorithm is applied to classify the graph data stream and output the graph data stream classification result. This includes: initializing traffic size threshold, duration threshold, and graph traversal depth threshold based on the graph data stream feature profiling model to obtain multi-dimensional determination parameters; configuring feature weights of the multi-dimensional determination parameters according to the network environment and application scenario; calculating a score for the graph data stream features after configuring feature weights to obtain a comprehensive determination score; classifying the graph data stream type according to the comprehensive determination score and preset determination rules to obtain a preliminary classification result; adding confidence information to the preliminary classification result and outputting the graph data stream classification result.

4. The method according to claim 1, characterized in that, Based on the graph data stream classification results, an optical circuit switching (OCS) path or an electrical packet switching (EPS) path is selected as the actual switching path for hybrid switching of the graph data streams using a switching path decision algorithm. This includes: monitoring the OCS and EPS resource status based on the graph data stream classification results and outputting a network resource status map; and, according to the network resource status map, directing the first graph data stream to the OCS optical path and the second graph data stream to the EPS channel, wherein the data volume of the first graph data stream is greater than the data volume of the second graph data stream.

5. The method according to claim 1, characterized in that, Collecting performance data to evaluate the actual exchange effect, and optimizing classification parameters and decision rules based on the performance data to update the classification parameters and decision rules, includes: collecting performance data after the actual exchange of the graph data stream in real time, determining the evaluation index results; establishing and training a machine learning model, inputting the evaluation index results into the machine learning model, and outputting an optimized parameter set; and generating the updated classification parameters and decision rules based on the optimized parameter set.

6. The method according to claim 1, characterized in that, Based on the updated classification parameters and decision rules, dynamic matching of graph data streams with network resources is performed to achieve intelligent transmission of graph data streams. This includes: formulating a graph data stream processing strategy based on the updated classification parameters and decision rules, wherein resources are reserved for high-priority graph data streams in the graph data stream processing strategy; adjusting the switching path configuration in real time based on the graph data stream processing strategy, performing batch configuration and pre-configuration, and outputting optimized configuration instructions; and performing dynamic matching of graph data streams with network resources based on the optimized configuration instructions to achieve intelligent transmission of graph data streams.

7. The method according to claim 1, characterized in that, Before performing dynamic matching of graph data streams and network resources, the method further includes: acquiring historical transmission records of the graph data streams, extracting traffic change patterns and resource occupancy characteristics, and fusing them to obtain a graph data stream state feature set; based on the graph data stream state feature set, constructing a Markov state transition model, and applying a reachability analysis algorithm to calculate the state transition probability to obtain traffic state prediction results at each time point; based on the traffic state prediction results and the existing network resource status, calculating the resource demand within a future time window; based on the resource demand within the future time window, performing path planning in conjunction with resource constraints of OCS and EPS to obtain a preset resource allocation strategy; and based on the preset resource allocation strategy, collaboratively optimizing the resource demands of multiple data streams to achieve dynamic matching of the graph data streams and network resources.

8. A hybrid exchange intelligent classification system based on graph data stream features, characterized in that, include: The acquisition module is used to acquire graph data streams, parse the graph data streams, extract multi-dimensional feature parameters including traffic size, duration, data packet interval and graph traversal depth, and construct a graph data stream feature profile model. The classification module is used to classify the graph data stream based on the graph data stream feature profiling model, apply a multi-dimensional feature threshold determination algorithm, and output the graph data stream classification result. The hybrid switching module is used to select either the Optical Circuit Switching (OCS) path or the Electrical Packet Switching (EPS) path as the actual switching path based on the graph data stream classification results and a switching path decision algorithm to perform hybrid switching of the graph data stream. The optimization module is used to collect performance data for evaluating the actual exchange effect, and to perform machine learning optimization on the classification parameters and decision rules based on the performance data to update the classification parameters and decision rules; the transmission module is used to perform dynamic matching of graph data streams with network resources according to the updated classification parameters and decision rules to realize intelligent transmission of graph data streams. The causal analysis module continuously monitors the transmission behavior of the graph data stream, acquires the transmission behavior of the graph data stream, and collects data on traffic changes, resource consumption, and performance indicators. Based on the traffic changes, resource consumption, and performance indicator data, it applies time series analysis methods to extract behavioral patterns and identifies the inter-flow influence relationships in the behavioral patterns through Granger causality tests, obtaining a preliminary causal association network. It uses the preliminary causal association network to calculate the influence propagation delay and propagation intensity, outputting a weighted time-series causal graph. It identifies key influence paths and propagation chains in the weighted time-series causal graph, obtaining a causal propagation model. Based on the causal propagation model, it predicts the chain reaction of decision adjustments to achieve optimal decision-making for multi-flow collaboration. The counterfactual decision optimization module acquires historical decision records of the graph data stream, extracts decision scenarios and effect data, and constructs a decision tree model to characterize decision rules based on the decision scenarios and effect data, obtaining a historical decision model. Based on the historical decision model, it generates multiple sets of hypothetical scenarios using the Monte Carlo method, generating a counterfactual scenario set. Based on the counterfactual scenario set, it performs simulation experiments to evaluate the effects of different decisions and outputs a comparison of decision effects. For each decision point, the optimal decision choice is identified from the simulation results of all counterfactual scenarios. Based on the optimal decision choice, the decision rules and evaluation criteria are updated to achieve continuous optimization of the decision-making mechanism.

Images