A dynamic advertisement frequency control optimization method based on multi-dimensional evaluation of family members

Abstract

The present application relates to the technical field of flow control, in particular to a dynamic advertisement frequency control optimization method based on multi-dimensional evaluation of family members, comprising obtaining device topology features and interaction logs to construct a heterogeneous mapping graph, and identifying a family unit by using graph representation learning and community discovery algorithm; by extracting behavior characteristics such as key Shannon entropy, time interval and response jitter, a classification model is used to determine and activate the active member identity; based on the user activity index and historical load data, a dynamic token bucket is initialized, and the load pressure coefficient is calculated by combining resource saturation and global residual water level weighting; the depth of the token bucket is dynamically adjusted by using a nonlinear negative feedback mapping mechanism to adjust the upper limit of frequency control in real time; cooperate with the flow interception hook and the double atomicity check mechanism to execute data distribution decision. Through identity perception and negative feedback regulation, the present application realizes adaptive optimization of quota allocation.

Description

Technical Field

[0001] This invention relates to the field of traffic control technology, specifically to a dynamic advertising frequency control optimization method based on multidimensional evaluation of family members. Background Technology

[0002] In the current IoT and smart home network environment, it is common for multiple users to share the same terminal device (such as smart TVs, tablets, etc.), a phenomenon known as the multi-user sharing of a single terminal interaction mode. However, existing multimedia information distribution technologies mainly rely on unique device identifiers for data push and traffic control. Since device IDs can only be mapped to physical hardware and cannot distinguish the specific individual currently operating the device, this many-to-one mapping relationship makes it difficult for the system to detect changes in the identity of the actual interacting entity, resulting in a series of technical defects at the data processing level.

[0003] First, due to the lack of effective identification of interactive entities, existing distribution systems often push a large number of high-load multimedia data packets (such as high-definition video streams or rich media materials) to non-target audiences (i.e., interactive entities with a very low probability of responding to the content). After receiving this invalid data, terminal devices still need to perform high-power operations such as decoding, rendering, and caching. This not only leads to the unnecessary consumption of terminal-side computing resources and storage space, but also causes a serious waste of network transmission bandwidth, increasing the risk of congestion in the overall network environment.

[0004] Secondly, in terms of traffic control and request scheduling, existing technologies typically employ static threshold control strategies based on the entire device (e.g., setting a daily limit on the total number of requests a device can receive). In multi-user shared scenarios, this coarse-grained control mechanism is highly susceptible to resource contention conflicts. For example, low-priority interactions may prematurely deplete the device's quota, causing subsequent high-priority critical interaction requests to be incorrectly intercepted or dropped by the system due to triggering device-level thresholds. This scheduling blocking problem severely reduces the effective throughput of the information distribution system and leads to delays in the system's response to high-value interaction requests.

[0005] Therefore, how to achieve dynamic flow control and fine-grained resource allocation in the complex interactive environment of shared devices is a technical problem that urgently needs to be solved in this field.

[0006] To address this, a dynamic advertising frequency control optimization method based on multidimensional assessment of family members is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a dynamic advertising frequency control optimization method based on multi-dimensional assessment of family members. Through identity perception and negative feedback adjustment, it achieves adaptive optimization of quota allocation. This includes acquiring device topology features and interaction logs to construct a heterogeneous mapping graph, and using graph representation learning and community discovery algorithms to identify family units; extracting behavioral features such as button Shannon entropy, time intervals, and response jitter, and using a classification model to determine and activate the identities of active members; initializing a dynamic token bucket based on user activity index and historical load data, and calculating the load pressure coefficient by weighting resource saturation and global remaining water level; dynamically adjusting the token bucket depth using a nonlinear negative feedback mapping mechanism to adjust the frequency control upper limit in real time; and executing data distribution decisions in conjunction with traffic interception hooks and a dual atomicity verification mechanism.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A dynamic advertising frequency control optimization method based on multidimensional assessment of family members includes:

[0010] Obtain the network topology characteristics of shared devices, account interaction logs, and advertising protocol parameters of advertising products; and obtain the total capacity of the family frequency control token pool.

[0011] A device-node mapping graph is constructed based on network topology features and account interaction logs; a family unit is determined using graph representation learning and community detection algorithms; and the corresponding currently active virtual interaction vertex is activated from the family unit based on device input commands using a node mapping model.

[0012] Obtain historical load data of the currently active virtual interaction vertex, calculate the interaction activity of the node based on the frequency and time span characteristics in the account interaction log; initialize a dynamic token bucket for the currently active virtual interaction vertex;

[0013] Based on historical load data and traffic burst tolerance in advertising protocol parameters, calculate the token bucket baseline capacity, calculate the token generation rate through interaction activity, calculate the resource consumption of the currently active virtual interaction vertex, calculate the global pressure coefficient based on the resource consumption and token pool capacity, and perform capacity scaling adjustment on the token bucket according to the global pressure coefficient to obtain the frequency control upper limit at the current moment.

[0014] When a data distribution request for an advertising product is received, check whether there are available tokens in the dynamic token bucket and the family total frequency control token pool and perform the operation.

[0015] Preferably, the process of obtaining the network topology characteristics of the shared devices, account interaction logs, and advertising protocol parameters of the advertising products, and obtaining the total family frequency control token pool capacity includes: collecting the media access control address hash value, wireless access point service set identifier, and received signal strength indication of the shared devices, and combining them to construct the network topology characteristics; capturing the sequence of human-machine interface device input instructions, application focus dwell time, and streaming media playback timestamp within a preset window, and generating the account interaction logs; extracting the network scheduling priority index and traffic burst tolerance from the advertising product metadata, and determining them as the advertising protocol parameters; parsing the maximum number of requests per unit time and the token reset period in the global traffic quota configuration file, and calculating their product as the total family frequency control token pool capacity.

[0016] Preferably, the process of constructing the device-node mapping graph includes: mapping the encrypted medium access control address in the network topology features to physical device vertices, and mapping the anonymized account unique identifier in the account interaction log to virtual interaction vertices; traversing all physical device vertices, and establishing an undirected spatial edge between two physical device vertices when their wireless access point service set identifiers are consistent; calculating the numerical deviation of the received signal strength indications between the two, and using a Gaussian kernel function to convert the numerical deviation into the signal environment similarity weight of the undirected spatial edge; traversing the account interaction log, and establishing a directed mapping edge between two physical device vertices when a specific virtual interaction vertex generates a human-machine interface device instruction on a specific physical device vertex, and using the instruction trigger frequency per unit time as the interaction strength weight of the directed mapping edge; combining the physical device vertices, virtual interaction vertices, undirected spatial edges, and directed mapping edges to generate a heterogeneous graph structure containing a node attribute matrix and a weighted adjacency matrix, which serves as the device-node mapping graph.

[0017] Preferably, the process of determining the family unit includes: defining a semantic meta-path from physical device vertex to virtual interaction vertex to physical device vertex in the device-node mapping graph, as a random walk path for constraint graph traversal; performing random walk sampling in the device-node mapping graph based on the semantic meta-path to generate a set of node sequences; training the set of node sequences using a skip word model with the goal of maximizing the co-occurrence probability of nodes within the sequence, mapping each vertex in the graph to a low-dimensional dense feature vector in a high-dimensional vector space; calculating the cosine similarity matrix between the low-dimensional dense feature vectors; constructing a weighted isomorphic graph based on the cosine similarity matrix; iteratively partitioning the weighted isomorphic graph using a modularity maximization algorithm until the modularity Q-value converges; and determining the set of vertices partitioned within the same module as the family unit.

[0018] Preferably, activating the corresponding currently active virtual interactive vertex includes: establishing a time sliding window of fixed length; serializing and sampling the device input command features within the sliding window; calculating the average time interval of key commands, the Shannon entropy of the key value probability distribution, and the jitter variance of the command response within the window; performing one-hot encoding on the current system time to generate time slice features; concatenating the average time interval, Shannon entropy, jitter variance, and time slice features into vectors to obtain a real-time operation feature vector; inputting the real-time operation feature vector into a pre-trained gradient boosting decision tree classification model to obtain the original discrimination score for each virtual interactive vertex; performing a normalized exponential function operation on the original discrimination score to calculate the posterior probability value; selecting the virtual interactive vertex with the highest posterior probability value that exceeds a preset activation threshold and marking its state as active.

[0019] Preferably, the process of initializing the dynamic token bucket for the currently active virtual interaction vertex includes: extracting the instruction timestamp sequence within the most recent preset period from the account interaction log, calculating the instruction arrival interval between adjacent instruction timestamps in the sequence, and taking the reciprocal of the instruction arrival interval as the instantaneous frequency feature; using an exponentially weighted moving average algorithm to iteratively smooth the instantaneous frequency feature to obtain a normalized interaction activity index; calculating the standard deviation of network throughput in the historical load data, multiplying the standard deviation, a preset burst multiplier coefficient, and the traffic burst tolerance in the advertising protocol parameters to determine the baseline burst capacity of the dynamic token bucket; using a linear affine transformation function to map the interaction activity index to the token generation rate of the dynamic token bucket; instantiating a new token bucket object, writing the baseline burst capacity into the bucket depth register of the current object, and writing the token generation rate into the token filling timer of the current object to complete the initialization of the dynamic token bucket.

[0020] Preferably, obtaining the personalized frequency control upper limit at the current moment includes: counting the actual number of tokens consumed by currently active virtual interaction vertices per unit time, calculating the ratio of the actual number of tokens consumed to the token generation rate, and determining it as the resource consumption saturation; querying the current available number of tokens in the total family frequency control token pool, calculating the ratio of the current available number of tokens to the capacity of the total family frequency control token pool, and determining it as the global remaining water level; dynamically configuring weight coefficients based on the network scheduling priority index in the advertising protocol parameters, using the weight coefficients to perform a weighted summation calculation of the resource consumption saturation and the complement of the global remaining water level, and obtaining the global load pressure coefficient; determining the negative correlation mapping curve according to the network scheduling priority index, mapping the global load pressure coefficient to a capacity scaling factor, multiplying the capacity scaling factor by the base bucket capacity, obtaining the adjusted dynamic bucket depth, and determining the adjusted dynamic bucket depth as the personalized frequency control upper limit at the current moment.

[0021] Preferably, the step of checking whether there are available tokens in the dynamic token bucket and the family total frequency control token pool and performing operations includes: deploying a traffic interception hook at the application layer of the network protocol stack, using the traffic interception hook to capture the Hypertext Transfer Protocol request data packet for the advertising product, and suspending the request data packet to the kernel-mode blocking wait queue; reading the current token count register value of the dynamic token bucket and the global quota semaphore value of the family total frequency control token pool, and using an atomic instruction to simultaneously determine whether the current token count register value and the global quota semaphore value are both greater than 0; if the determination result is yes, then performing an atomic decrement operation to simultaneously deduct one unit of token count and global quota, releasing the request data packet from the blocking wait queue, and writing the request data packet into the send buffer of the network interface controller for physical transmission; if the determination result is no, then directly removing the request data packet from the blocking wait queue, constructing a rejection response message containing an excessive request status code, and returning the rejection response message to the initiator of the request data packet.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. By utilizing semantic meta-path and modularity maximization algorithms, high-confidence family units can be automatically clustered in open network environments without explicit user relationship data. By combining Shannon entropy of key values ​​and instruction response jitter variance, the operating habits of users of different age groups can be distinguished without collecting camera images or specific content logs. This non-intrusive recognition mechanism solves the technical problem of difficulty in defining the main user of shared devices in privacy-restricted scenarios, and significantly improves the matching accuracy and discriminability of frequency control strategies for specific user behavior patterns.

[0024] 2. By using an exponential decay mechanism to assign higher weight to recent data, the token generation rate can dynamically follow changes in the user's current activity level, effectively solving the parameter lag problem in the cold start phase; by using the standard deviation to quantify the physical fluctuation range of traffic to set the baseline burst capacity, the buffer size can cover the actual signal jitter range, ensuring the availability of basic services while reducing the burst traffic packet loss rate caused by parameter mismatch.

[0025] 3. Mapping the global load pressure coefficient to a capacity scaling factor can proactively tighten the global resource supply when network quality deteriorates, preventing system overload. The inverse variance weighted algorithm is used to fuse local demand deviation and global supply deviation, and calculate the capacity scaling factor. This allows the system to not only smoothly expand when resources are plentiful, but also trigger cascading shrinkage when resources are contested, maximizing the utilization efficiency of limited network resources and the overall throughput stability of the system. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a dynamic advertising frequency control optimization method based on multidimensional assessment of family members according to the present invention;

[0027] Figure 2 This is a schematic diagram of the activation process of the corresponding currently active virtual interactive vertex according to the present invention;

[0028] Figure 3 This is a schematic diagram of the process for obtaining the personalized frequency control upper limit at the current moment according to the present invention. Detailed Implementation

[0029] 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.

[0030] Please see Figures 1 to 3 This invention provides a dynamic advertising frequency control optimization method based on multi-dimensional evaluation of family members, the technical solution of which is as follows:

[0031] Example 1:

[0032] A dynamic advertising frequency control optimization method based on multidimensional assessment of family members, specifically as follows: Figure 1 As shown, it includes:

[0033] Obtain the network topology characteristics, account interaction logs, and advertising protocol parameters of shared devices, and obtain the total capacity of the family frequency control token pool;

[0034] A device-node mapping graph is constructed based on network topology features and account interaction logs; a family unit is determined using graph representation learning and community detection algorithms; and the corresponding currently active virtual interaction vertex is activated from the family unit based on device input commands using a node mapping model.

[0035] Obtain historical load data of the currently active virtual interaction vertex, calculate the interaction activity of the node based on the frequency and time span characteristics in the account interaction log; initialize a dynamic token bucket for the currently active virtual interaction vertex;

[0036] Based on historical load data and traffic burst tolerance in advertising protocol parameters, calculate the token bucket baseline capacity, calculate the token generation rate through interaction activity, calculate the resource consumption of the currently active virtual interaction vertex, and calculate the global pressure coefficient based on the resource consumption and token pool capacity; perform capacity scaling adjustment on the token bucket according to the global pressure coefficient and the network scheduling priority index of advertising protocol parameters to obtain the frequency control upper limit at the current moment.

[0037] When a data distribution request for an advertising product is received, check whether there are available tokens in the dynamic token bucket and the family total frequency control token pool and perform the operation.

[0038] Furthermore, the process of obtaining the network topology characteristics of the shared devices, account interaction logs, and advertising protocol parameters of the advertising products, and obtaining the total capacity of the family frequency control token pool, includes: collecting the media access control address hash value, wireless access point service set identifier, and received signal strength indication of the shared devices, and combining them to construct the network topology characteristics; capturing the sequence of input instructions from the human-machine interface device within a preset window, the application focus dwell time, and the streaming media playback timestamp, and generating the account interaction logs; extracting the network scheduling priority index and traffic burst tolerance from the advertising product metadata, and determining them as the advertising protocol parameters; parsing the maximum number of requests per unit time and the token reset period in the global traffic quota configuration file, and calculating their product as the total capacity of the family frequency control token pool.

[0039] Specifically, the underlying network driver interface of the shared device is invoked to read the Media Access Control (MAC) address of the current physical network card. To ensure data privacy, the plaintext address is not stored directly. Instead, a preset random salt value is introduced, and the MAC address is hashed using the SHA-256 encryption algorithm. The first 16 hexadecimal characters of the hash result are extracted as the device's unique hardware fingerprint. Next, the current wireless network environment is scanned to read the physical address of the currently connected wireless access point and the received signal strength indicator (RSSI) value, for example, if the RSI is -65dBm. Subsequently, the hashed MAC address, the physical address of the wireless access point, and the RSI value are sequentially encapsulated and combined to construct a three-dimensional feature vector, which serves as the network topology feature representing the physical location of the device.

[0040] Meanwhile, within a preset time window, by acquiring interrupt events from the input subsystem, the system captures in real-time the sequence of human-machine interface input commands issued by the user via remote control or external devices, focusing on recording key value codes including the confirmation key, back key, and directional keys. Simultaneously, the system monitors the operating system's window manager state, calculates the time difference between foreground application package name switching, and determines the duration of current application focus. Furthermore, when a streaming media playback process is detected to be active, the system timestamp when the user triggers a play or pause action is recorded. The aforementioned command sequences, duration data, and timestamp information are packaged and serialized into JSON format log data, thereby generating an account interaction log.

[0041] While acquiring the aforementioned basic data, the metadata header file returned with the ad initialization request is parsed. The network scheduling priority index value and traffic burst tolerance coefficient, which represent business priorities, are directly extracted from the extended fields of this file. After these parameters are extracted, they are temporarily stored in the memory configuration table for subsequent definition of the service baseline of this ad distribution request, ensuring that ad products with different priorities can be matched with corresponding network resources.

[0042] Finally, the global traffic quota configuration file stored in the cloud is read, and the maximum number of requests allowed per unit time (e.g., per minute) and the token bucket reset cycle parameter defined in the file are parsed. A multiplication operation is performed in the local processing unit, multiplying the maximum number of requests per unit time by the token reset cycle. The result is the physical capacity limit of the total family frequency control token pool. Based on this calculation result, a global semaphore object is initialized in the shared memory area, and the calculation result is assigned to the semaphore's initial counter, thus completing the acquisition of the total family frequency control token pool capacity.

[0043] By using hash algorithms to de-identify hardware identifiers, the physical fingerprint of the device is established while ensuring privacy. HID commands and focus states are captured directly from the kernel layer, avoiding the risk of application layer data forgery and ensuring the objectivity of behavioral data. At the same time, SLA parameters are parsed and the total family frequency control token pool capacity is calculated, transforming the abstract business protocol into specific network resource quotas, providing accurate global total constraints and initialization benchmarks for subsequent traffic shaping and dynamic allocation.

[0044] Furthermore, the process of constructing the device-node mapping graph includes: mapping the encrypted medium access control address in the network topology features to physical device vertices, and mapping the anonymized account unique identifier in the account interaction log to virtual interaction vertices; traversing all physical device vertices, and establishing an undirected spatial edge between two physical device vertices when their wireless access point service set identifiers are consistent; calculating the numerical deviation of the received signal strength indications between the two, and using a Gaussian kernel function to convert the numerical deviation into the signal environment similarity weight of the undirected spatial edge; traversing the account interaction log, and establishing a directed mapping edge between two physical device vertices when a specific virtual interaction vertex generates a human-machine interface device instruction on a specific physical device vertex, and using the instruction trigger frequency per unit time as the interaction strength weight of the directed mapping edge; combining the physical device vertices, virtual interaction vertices, undirected spatial edges, and directed mapping edges to generate a heterogeneous graph structure containing a node attribute matrix and a weighted adjacency matrix, which serves as the device-node mapping graph.

[0045] Specifically, first, an empty heterogeneous graph data object is initialized in memory. The network topology feature set is read, and the encrypted medium access control address is extracted. A physical device vertex object is instantiated for each independent encrypted MAC address, and the corresponding device type is written as an attribute to the attribute list of that vertex. Next, the anonymous account unique identifier ID in the account interaction log is parsed, and a virtual interaction vertex object is instantiated for each independent account ID. A global vertex index table is maintained to ensure that the IDs of all vertices in the graph are unique.

[0046] After vertex instantiation is completed, the construction process of undirected spatial edges is executed. All physical device vertices are traversed, and vertices with the same wireless access point service set identifier are grouped into the same cluster using the hash bucket algorithm. Within the same cluster, any two physical device vertices are paired up, and the received signal strength indication values ​​recorded in the network topology features are read. The absolute difference between the two received signal strength indication values ​​is calculated. In order to accurately quantify the similarity of physical distance, the Gaussian radial basis kernel function is used to perform nonlinear mapping calculation on the difference. The specific calculation logic is to calculate the exponent of the natural constant e, and divide the square of the received signal strength indication difference with a negative exponent by twice the square of the preset bandwidth parameter. The normalized value (range 0 to 1) obtained by the calculation is used as the signal environment similarity weight. An undirected spatial edge with this weight is established between the corresponding two physical device vertices, thereby logically connecting devices with close physical distances. The method of converting the numerical deviation into signal environment similarity weights for the undirected spatial edges using a Gaussian radial basis function specifically includes setting the bandwidth parameter of the kernel function to 10. This value is set according to the signal standard deviation in the actual home environment, which can ensure that signal differences within 10dB are mapped to a high similarity. When the signal deviation between two devices is closer to 0, the output similarity weight is closer to 1. As the deviation increases, the weight decays exponentially. This nonlinear mapping mechanism can effectively widen the numerical difference between devices in the same room and devices across rooms, thereby providing highly discriminative spatial features for the subsequent accurate clustering of family units.

[0047] Subsequently, the process of constructing directed mapping edges is executed, traversing the account interaction log stream to identify the operation records of specific virtual interaction vertices on specific physical device vertices. Using minutes as the time granularity, the total number of human-machine interface device commands (including clicks, swipes, confirmations, etc.) generated by the virtual account on the physical device is counted. This number is divided by the time window length to obtain the command trigger frequency. When the frequency is greater than 0, a directed mapping edge is established from the virtual interaction vertex to the physical device vertex, and the command trigger frequency is directly assigned as the interaction strength weight of the directed edge, thereby characterizing the stickiness of human control over objects.

[0048] Finally, the generated vertex and edge structure is subjected to matrix serialization. Following the order of the global vertex index table, a sparse weighted adjacency matrix of dimension N×N is constructed, where matrix element A... ij Store the edge weights between vertex i and vertex j (0 if there are no edges); at the same time, construct a node attribute matrix of dimension N×F (F is the feature dimension) to store the category label and original features of each vertex. Encapsulate the weighted adjacency matrix and the node attribute matrix into a standard heterogeneous graph data package as the final generated device-node mapping graph.

[0049] Before calculating the numerical deviation between the two received signal strength indicators, the method further includes: establishing a Kalman filter model for the physical device vertex; inputting the collected real-time received signal strength indicator values ​​as observation values ​​into the Kalman filter model; pre-setting the process noise covariance and measurement noise covariance according to the device type of the physical device vertex; iteratively calculating through two stages of prediction and update to output a smoothed signal strength estimate; and using the smoothed signal strength estimate to calculate the numerical deviation between the two.

[0050] In constructing the device-node mapping map, to address the issue that received signal strength indicators are susceptible to multipath effects, shadow fading, and environmental thermal noise interference in wireless propagation environments, this embodiment introduces a signal smoothing mechanism based on Kalman filtering before calculating numerical deviations. First, a discrete-time state-space model is established for each physical device vertex, and the raw intensity values ​​collected at each moment are regarded as the observation values ​​input to the model. Based on the device attributes of the physical device vertex, the process noise covariance and measurement noise covariance are preset: for fixed devices such as smart TVs, a smaller process noise covariance is preset to enhance the confidence in historical estimates; for mobile devices such as mobile phones or tablets, a larger process noise covariance is preset to quickly track signal reference drift caused by position changes.

[0051] During the iterative calculation process, the prediction phase and the update phase are executed in a loop. In the prediction phase, the prior state at the current moment is calculated using the optimal estimate from the previous moment. In the update phase, the prior estimate is corrected by calculating the gain ratio between the observed value and the predicted value and using the real-time acquired signal values, thereby outputting a smoothed signal strength estimate. This estimate effectively filters out instantaneous signal spikes caused by people walking around or electrical interference. Subsequently, the smoothed estimate is used to calculate the numerical deviation between different physical devices, and the Gaussian kernel function is used to convert it into a signal environment similarity weight, ensuring that the spatial edges in the map can truly reflect the physical distance relationship between devices, providing high-confidence underlying topological data for subsequent identification of household units.

[0052] When identifying family units, a dual anchoring strategy of physical topology features and behavioral association features is adopted. Although the semantic meta-path establishes a logical connection through shared accounts, a signal strength similarity weight based on Kalman filtering smoothing is introduced simultaneously. For devices that do not share accounts but are physically close (such as in the same living room), even if the meta-path logic is weak, the signal environment similarity of their undirected spatial edges can still remain high. This design ensures the discovery of family members and also eliminates edge devices with occasional connections through the consistency of physical space. Thus, in an open network without explicit user relationship data, high-confidence family unit entities are automatically clustered.

[0053] By utilizing the Gaussian kernel function to process the deviation of the wireless signal strength indicator signal, the discrete and jittery physical signal is transformed into a continuous and normalized spatial similarity weight, effectively suppressing environmental noise caused by multipath effects. At the same time, a directed mapping edge is established based on the frequency of HID commands, which accurately quantifies the user's control stickiness to specific devices. This heterogeneous graph that integrates physical topology and behavioral logic fully preserves the multidimensional semantic information in the network, providing a high-fidelity topological data foundation for accurate identification of family and community structures in the absence of explicit social relationship data.

[0054] Further, the process of determining the family unit includes: defining a semantic meta-path from physical device vertex to virtual interaction vertex to physical device vertex in the device-node mapping graph, as a random walk path for constraint graph traversal; performing random walk sampling in the device-node mapping graph based on the semantic meta-path to generate a set of node sequences; training the set of node sequences using a skip word model with the goal of maximizing the co-occurrence probability of nodes within the sequence, mapping each vertex in the graph to a low-dimensional dense feature vector in a high-dimensional vector space; calculating the cosine similarity matrix between the low-dimensional dense feature vectors; constructing a weighted isomorphic graph based on the cosine similarity matrix; iteratively partitioning the weighted isomorphic graph using a modularity maximization algorithm until the modularity Q-value converges; and determining the set of vertices partitioned within the same module as the family unit.

[0055] Specifically, after constructing the device-node mapping graph, a graph-based feature learning process is first executed under the control of the processor. Semantic meta-path patterns of physical device vertices-virtual interaction vertices-physical device vertices are defined in memory, serving as the path rules for constraining graph traversal. In this embodiment, the number of random walks initiated for each vertex in the graph is set to 10, and the step size of a single walk is set to 40 nodes. This step size parameter is set because in the network topology of a home LAN, the interaction link radius between devices and users is usually limited. A step length of 40 nodes is sufficient to cover a typical home network structure including smart TVs, mobile terminals, and IoT devices, while preventing the walker from excessively drifting to edge interference nodes caused by signal noise. Setting the number of walks to 10 is to control the memory overhead and time cost of sampling calculations while ensuring coverage of sparse regions of the graph, ensuring that the system can complete the corpus generation within milliseconds.

[0056] Next, the generated node sequence corpus is input into a pre-defined skip character model for vectorized training. In this embodiment, the model is optimized using negative sampling, with the negative sampling number set to 15 to improve computational efficiency without sacrificing accuracy. Simultaneously, a hierarchical normalized exponential function is used as the target optimization direction for the output layer. The feature vector dimension of the embedding layer is initialized to 128 dimensions, and the context window size is set to 5. The 128-dimensional dimension is chosen because it conforms to the characteristics of computer memory page alignment, which is beneficial for performing parallel matrix operations. Furthermore, this dimension is sufficient to preserve the complex semantic differences between heterogeneous nodes (physical devices and virtual accounts) in a low-dimensional space, avoiding information loss due to excessively low dimensions or overfitting due to excessively high dimensions. The window size is set to 5 because this length precisely covers the device... The second-order proximity relationship between account and device enables the model to effectively capture the indirect association features generated by sharing virtual accounts between different physical devices within the same household. After obtaining the feature vectors of all nodes, the vector operation unit is called to calculate the cosine similarity between all node pairs. A similarity filtering threshold of 0.5 is set, and a hard threshold truncation operation is performed to retain only node pairs with similarity greater than the threshold. The similarity value is directly used as the new edge weight to construct a weighted isomorphic graph. Setting 0.5 as the threshold (corresponding to a vector angle of 60 degrees) is to filter out weak connections with only weak associations or occasional interactions, i.e., to remove noise interference and retain only the core backbone connections with strong structural equivalence. This significantly reduces the size of the adjacency matrix of the subsequent community partitioning algorithm and improves the purity and efficiency of clustering calculation.

[0057] In this embodiment, the number of iterations of the jump character model is set to 100 rounds, and the convergence criterion is: when the loss function decreases by less than 0.01 in 5 consecutive iterations, it is determined that the model has captured stable node structure features. At this time, training is stopped and 128 low-dimensional dense feature vectors of all vertices are output.

[0058] Finally, the modularity maximization algorithm (Louvain's algorithm in this embodiment) is run on the constructed weighted isomorphic graph for iterative partitioning. During algorithm initialization, each node is marked as an independent community, and then iterative optimization is performed. In this embodiment, the convergence criterion is that the modularity Q-value increment is less than 0.001 for three consecutive iterations. This convergence threshold is set to balance the algorithm's running time and partitioning accuracy. When the Q-value increment is lower than this value, it indicates that the partitioning of the community structure has become stable, and the marginal benefit of continuing iteration is extremely low. At this time, the calculation is stopped, and all physical device vertices and virtual interaction vertices that are finally partitioned into the same community module are locked into a set, and this set is confirmed as an independent family unit, completing the mapping from network topology to family entity.

[0059] By using graph representation learning to transform heterogeneous physical devices and virtual accounts into low-dimensional dense vectors, the implicit second-order proximity relationships and structural equivalence between nodes are effectively captured. Combined with the modularity maximization algorithm to iteratively partition the weighted isomorphic graph, it can automatically converge to the globally optimal community structure from a mathematical optimization perspective, avoiding the subjective bias of manually set rules. This accurately eliminates edge noise devices with weak connections, establishing a high-confidence range of the main family unit for the frequency control system.

[0060] Further, activating the corresponding currently active virtual interactive vertex includes: establishing a time sliding window of fixed length; serializing and sampling the device input command features within the sliding window; calculating the average time interval of key commands, the Shannon entropy of the key value probability distribution, and the jitter variance of the command response within the window; performing one-hot encoding on the current system time to generate time slice features; concatenating the average time interval, Shannon entropy, jitter variance, and time slice features into vectors to obtain a real-time operation feature vector; inputting the real-time operation feature vector into a pre-trained gradient boosting decision tree classification model to obtain the original discrimination score for each virtual interactive vertex; performing a normalized exponential function operation on the original discrimination score to calculate the posterior probability value; selecting the virtual interactive vertex with the highest posterior probability value that exceeds a preset activation threshold and marking its state as active. The specific process is as follows: Figure 2 As shown.

[0061] Specifically, after determining the family unit, a real-time activation detection process for virtual interaction vertices is initiated. First, a fixed-length circular buffer is allocated in memory, and a time sliding window with a time span of 60 seconds is established. Within this window, the system serializes and samples the human-machine interface device commands reported by the underlying input subsystem. In this embodiment, the sampling limit is set to the 50 most recent key commands within the window. If there are insufficient commands, zeros are added. The command sequence is traversed, and the difference between the timestamps of adjacent commands is calculated. The arithmetic mean is taken to obtain the average time interval, which represents the operation rate. The frequency of different key values ​​(such as up, down, left, right, confirm, return) in the sequence is statistically analyzed, and the Shannon entropy of the probability distribution of key values ​​is calculated using the Shannon formula, which represents the randomness and disorder of the operation. At the same time, the dispersion of the command response delay relative to the average delay, i.e., the jitter variance, is calculated, which represents the muscle control stability of the user's finger movements (for example, the jitter variance of elderly users is usually greater than that of young users).

[0062] While extracting behavioral features, the current system clock is read, the hour value (0-23) is extracted, and one-hot encoding is performed on it. In this embodiment, a 24-dimensional sparse vector is constructed, the index dimension corresponding to the current hour is set to 1, and the other dimensions are set to 0, thereby generating time slice features. Subsequently, a vector concatenation operation is performed, concatenating the average time interval (1-dimensional), the Shannon entropy of the key value probability distribution (1-dimensional), the jitter variance of the instruction response (1-dimensional) obtained above with the time slice features (24-dimensional), and combining them to generate a real-time operation feature vector with a total dimension of 27. This vector comprehensively represents the user's operation frequency features, behavioral randomness features, physiological control stability features, and temporal distribution context information.

[0063] Next, the pre-trained gradient boosting decision tree classification model loader is invoked. This model has been trained offline using historical operation data with user labels and contains an ensemble predictor composed of multiple decision trees. The system feeds the model with the constructed 27-dimensional real-time operation feature vector as input data. The model traverses the internal decision paths in parallel and finally outputs the raw discrimination score corresponding to each virtual interaction vertex (such as father, mother, and child) in the output layer. The raw score is an unnormalized real value that reflects the model's initial confidence in the identity of each user.

[0064] In this embodiment, the gradient boosting decision tree classification model is specifically constructed using the XGBoost algorithm framework. Compared to the basic decision tree, this model can effectively handle nonlinear features and noisy data in key press behavior. The real-time operation feature vector consists of 27 independent dimensions to ensure that the model can capture user features from multiple perspectives.

[0065] Behavioral timing characteristics (1D): The average time interval of the first 50 key presses within the sliding window, reflecting the user's operation speed.

[0066] Behavioral randomness characteristics (1-dimensional): The probability distribution entropy of key values ​​calculated using Shannon entropy reflects the degree of dispersion of the operation logic.

[0067] Stability characteristics (1-dimensional): Jitter variance of command response, used to quantify the delay fluctuations caused by differences in muscle control among users of different ages.

[0068] Temporal context features (24 dimensions): One-hot encoding is performed on the entire 24 hours, with each dimension corresponding to one hour, to capture the operational preferences of different family members at different times.

[0069] During the offline training phase, labeled interaction data of each member within the family unit is collected for no less than 30 days. The sample size ensures that each member has no less than 100 positive samples. The learning rate is set between 0.05 and 0.15, with 0.1 being preferred in this embodiment to balance convergence speed and generalization ability. The tree depth is set to 4 to 8 layers, with 6 layers in this embodiment to prevent overfitting. The number of iteration rounds is set to 100 rounds. When the improvement in the accuracy of the validation set is less than 0.1% for 5 consecutive iterations and the overall accuracy is stable above 95%, the model is considered to have converged and training is stopped. After inputting real-time features into the model, the model outputs the original discrimination score of each virtual interaction vertex. The scores are mapped to a posterior probability distribution with a sum of 1 using a normalized exponential function (Softmax logic) to determine the identity confidence level.

[0070] To cope with changes in family members or the evolution of users' long-term behavior (such as changes in operating habits due to the growth of teenagers), it has the ability to learn incrementally. When a new member is added or the recognition confidence is consistently abnormal, it will automatically extract the 50 most recently generated feature samples and fine-tune the weights of the existing model with a low learning rate (0.01 in this embodiment) to ensure the continuity of recognition accuracy.

[0071] Finally, a normalized exponential function operation is performed on the output raw discrimination scores. Specifically, the natural exponent of each score is calculated and divided by the sum of the exponents of all scores, thereby mapping the raw scores to a posterior probability distribution with a sum of 1. This probability distribution is scanned, and the virtual interaction vertex with the highest probability value is selected as a candidate. This highest probability value is compared with a preset activation threshold. In this embodiment, the activation threshold is set to 0.85. If the highest probability value exceeds this threshold, it indicates that the system has a very high degree of confirmation of the current operator's identity. The status flag register of the virtual interaction vertex is immediately set to "active state," and subsequent frequency control strategies are configured based on the identity of this node. If it is lower than this threshold, the current state remains unchanged or it is marked as "unknown user" to prevent policy errors caused by misjudgment.

[0072] To address the potential fluctuations in behavioral characteristics caused by different interactive tasks, the average key press interval and response jitter variance of users are largely limited by individual finger motor skills and physiological reaction habits. These characteristics exhibit strong cross-task consistency across different application interfaces (such as search interfaces or content selection interfaces). In terms of feature dimension processing, by performing serial sampling and normalization exponential processing on the features within the sliding window, combined with the regularization term built into the XGBoost algorithm, common fingerprints in behavioral features can be effectively extracted and noise suppressed. When task interference causes ambiguity in the posterior probability (i.e., the difference between the first and second high probabilities is too small), second-order conflict discrimination is automatically triggered, and the behavioral prior of historical time slices is introduced for correction, thereby avoiding the risk of misidentification caused by differences in behavior in a single task.

[0073] The activation of the corresponding currently active virtual interaction vertex further includes: when the difference between the first high probability value and the second high probability value in the calculated posterior probability values ​​is less than a preset fuzzy confidence interval threshold, triggering a second-order conflict discrimination process; retrieving the historical active weight distribution corresponding to the current time slice in the historical account interaction log, performing element-level weighted fusion of the historical active weight distribution and the original discrimination score, and recalculating the posterior probability; if the weighted posterior probability value still does not exceed the preset activation threshold, then marking the state of the currently active virtual interaction vertex as a restricted activation state, and loading a preset family shared basic frequency control template.

[0074] Specifically, after the gradient boosting decision tree classification model outputs the posterior probability distribution of each vertex, the processor monitors the discriminative power of the probability distribution in real time. If the difference between the calculated first and second highest probability values ​​is less than the preset fuzzy confidence interval threshold (set to 0.05 in this embodiment), it is determined that the current operation features are insufficient to uniquely identify the user. A second-order conflict discrimination process is then triggered. During the discrimination process, historical account interaction logs stored locally or in the cloud are retrieved. Based on the current system time, the historical active weight distribution within the corresponding time slice is extracted. This distribution reflects the statistical prior probability of different family members using the device during specific time periods (such as evening rush hour or early morning). Subsequently, element-level weighted fusion is performed, multiplying the original discrimination score of each virtual interaction vertex with its corresponding historical active weight. The algorithm calculates and performs a normalized exponential function operation on the fused result to recalculate the posterior probability. After the second calculation, the posterior probability value is scanned again. If the corrected highest posterior probability value still does not exceed the preset activation threshold (e.g., 0.85), it is determined that the current state is in a high-uncertainty interaction state. In this state, a specific individual identity will not be forcibly activated. Instead, the state of the currently active virtual interaction vertex will be marked as a restricted activation state. In this state, when the distribution decision is executed through the traffic interception hook, the personalized frequency control limit of a specific member will no longer be called. Instead, a preset family-shared basic frequency control template will be loaded. This template usually adopts a conservative traffic quota and a high request interval limit to prioritize the security of the family's total frequency control token pool and the stability of the overall system load in the case of unknown identity. Through this implementation method, in edge scenarios with highly overlapping behavioral characteristics, prior information in the time dimension can be used to resolve conflicts. Combined with the restricted activation mechanism, it avoids excessive resource allocation caused by misjudgment, which significantly improves the robustness of the flow control strategy in complex shared scenarios.

[0075] In this embodiment, the acquisition and application process of the historical activity weights is as follows: By traversing the account interaction logs within a preset period, the entire 24 hours are divided into multiple time slices of fixed length, and the frequency of human-computer interface device commands generated by each virtual interaction vertex within the family unit is counted for each time slice; by calculating the proportion of the command frequency of a specific member in the corresponding time slice to the total frequency of the family, and using an exponentially weighted moving average algorithm to smooth and iterate the values ​​over multiple days, a historical activity weight distribution matrix representing the temporal behavioral preferences of members is generated; when the difference between the first high probability value and the second high probability value in the posterior probability values ​​output by the gradient boosting decision tree is less than a preset fuzzy confidence interval threshold, If the current operation features are found to have overlapping identities, a second-order conflict discrimination process is triggered. The historical active weights corresponding to the current time slice are retrieved from the memory and then element-wise weighted and fused with the original discrimination scores output by the classification model. This allows for the recalculation of the posterior probability values ​​of each vertex using prior information about historical behavior. If the highest posterior probability value after weighted correction still does not exceed the preset activation threshold, it indicates that the current interaction state has extremely high identity ambiguity. The state of the currently active virtual interaction vertex is marked as a restricted activation state, and a family-shared basic frequency control template containing conservative traffic quotas and high latency tolerance is loaded. This prioritizes the security of the family's total frequency control token pool and the stability of the overall system throughput in the case of unclear identities.

[0076] By establishing a time-sliding window sampling mechanism and calculating the average interval, Shannon entropy, and jitter variance of key presses, the system can quantify biodynamic characteristics reflecting user operation frequency, behavioral randomness, and physiological control stability from the input signal. Combined with time-slice one-hot encoding, a real-time operation feature vector is constructed, and the nonlinear fitting capability of the gradient-enhanced decision tree model is used for probabilistic inference. This solves the technical problem of difficulty in real-time differentiation of the current operator in shared device scenarios without explicit login. This activation mechanism based on Softmax posterior probability ensures that the system can lock onto family members with specific operating habits within milliseconds, providing accurate dynamic identity criteria for subsequent loading of personalized frequency control strategies.

[0077] Further, the process of initializing the dynamic token bucket for the currently active virtual interaction vertex includes: extracting the instruction timestamp sequence within the most recent preset period from the account interaction log, calculating the instruction arrival interval between adjacent instruction timestamps in the sequence, and taking the reciprocal of the instruction arrival interval as the instantaneous frequency feature; using an exponentially weighted moving average algorithm to iteratively smooth the instantaneous frequency feature to obtain a normalized interaction activity index; calculating the standard deviation of network throughput in the historical load data, multiplying the standard deviation, a preset burst multiplier coefficient, and the traffic burst tolerance in the advertising protocol parameters to determine the baseline burst capacity of the dynamic token bucket; using a linear affine transformation function to map the interaction activity index to the token generation rate of the dynamic token bucket; instantiating a new token bucket object, writing the baseline burst capacity into the bucket depth register of the current object, and writing the token generation rate into the token filling timer of the current object to complete the initialization of the dynamic token bucket.

[0078] First, the instruction timestamp sequence within the most recent 300 seconds (i.e., a preset period) is extracted from the account interaction log. This sequence is then iterated through, and the difference (in milliseconds) between adjacent instruction timestamps is calculated to obtain the instruction arrival interval sequence. To quantify the instantaneous operation frequency, a reciprocal operation is performed, and the reciprocal of each instruction arrival interval is determined as the instantaneous frequency feature. To eliminate random fluctuations in operation frequency, an exponentially weighted moving average algorithm is used to iteratively smooth the instantaneous frequency feature. In this embodiment, a smoothing coefficient is set. The weight is 0.2, meaning the current frequency accounts for 20% of the weight, while the historical average accounts for 80%. The calculation formula is as follows:

[0079] ;

[0080] in Indicates the instantaneous frequency at the current moment. This represents the smoothed value from the previous time step. This represents the smoothed value at the current moment. This represents the smoothing coefficient.

[0081] Subsequently, the system calculates the token generation rate using a linear affine transformation function ( To establish a mapping relationship, in this embodiment, a linear slope is set. The base intercept is 10. The value is 5 (representing the minimum guaranteed rate). This parameter can be adjusted based on different average package sizes, using the previously calculated interaction activity index as a variable. Substituting the values ​​into the function yields the token generation rate of the dynamic token bucket (unit: tokens / second), ensuring that the token supply speed is proportional to the intensity of real-time user interaction.

[0082] Through iterative calculations, a stable normalized interaction activity index is obtained. This index can keenly reflect the user's current activity trend and filter out occasional accidental touch noise. Specifically, the normalization process performs min-max normalization, maintaining the user's minimum and maximum activity values ​​in historical interaction records in real time, and mapping the current smoothed value to the range of 0 to 1. This normalization process ensures that users with different interaction styles can be evaluated under the same feature scale, providing standardized data input for the subsequent linear mapping of token generation rate. Simultaneously, historical load data of the node over the past 24 hours is retrieved, and a network throughput sample sequence with a sampling period of minutes is extracted. First, the arithmetic mean of all sample values ​​in the sequence is calculated. Then, the square of the difference between each sample value and the mean is calculated, and all these squares are summed. The sum of the squares is divided by the total number of samples to obtain the variance. The square root of the variance is then taken to obtain the standard deviation of the network throughput. The original value of the traffic burst tolerance in the advertising protocol parameters is extracted. This original value is usually a value between 0 and 100. It is mapped to a basic elasticity range through a preset normalization ratio. In this embodiment, the original value is converted into a basic adjustment score between 1.0 and 1.5 by dividing it by 200 and adding 1.0. The network scheduling priority index in the advertising protocol parameters is further retrieved, and a hierarchical compensation table is preset.

[0083] First Tier (High Priority): When the network scheduling priority index is between level 1 and 3 (the smaller the value, the higher the priority), it is determined to be a high service quality requirement. At this time, a positive compensation increment is added on top of the basic adjustment score. This increment is set to a fixed value between 0.2 and 0.3, which is 0.2 in this embodiment.

[0084] Second tier (medium priority): When the network scheduling priority index is between level 4 and 6, the cumulative compensation increment is set to 0.1.

[0085] Third tier (normal priority): When the network scheduling priority index is greater than 6 levels, the compensation increment is set to 0, that is, the basic adjustment score is directly used as the final coefficient.

[0086] The above-mentioned basic adjustment score is added to the corresponding compensation increment to obtain the final normalized tolerance adjustment coefficient. Then, a compound multiplication operation is performed to multiply the calculated standard deviation, the preset burst multiplier coefficient (2.0 in this embodiment), and the finally determined normalized tolerance adjustment coefficient. The result of the above compound multiplication operation is rounded up to formally determine the baseline burst capacity of the dynamic token bucket. Next, a new token bucket object is instantiated in the memory heap area. Through low-level memory pointer operations, the calculated baseline burst capacity value is written to the bucket depth register of the object in real time, and the calculated token generation rate value is written to the token filling timer of the object. The timer periodically adds tokens to the bucket according to the countdown of the time at the rate, completing the parameter adaptive dynamic token bucket initialization.

[0087] Although the calculation process is based on standard deviation, by introducing a traffic burst tolerance and a preset burst multiplier for composite weighting, it means that the depth of the token bucket does not only cover the normal fluctuations in a statistical sense, but also reserves sufficient physical buffers for extreme burst scenarios through the adjustment of the multiplier. For advertising products with strong burst characteristics such as video streams, the network scheduling priority index in the metadata is extracted. For high-priority services, positive compensation increments are automatically added on top of the basic capacity, thereby ensuring that the token bucket depth can cover the heavy-tail traffic fluctuations in the real network environment at peak traffic moments, effectively avoiding the problem of false rate limiting and packet loss caused by excessively narrow parameter settings.

[0088] By extracting the inverse of the instruction interval and applying an exponentially weighted moving average algorithm, operational noise is effectively filtered out while behavioral trends are preserved. A baseline burst capacity is set based on the standard deviation of network throughput, and statistical principles are used to ensure that the buffer size can cover the fluctuation range of historical traffic, avoiding false flow limiting due to insufficient capacity. A positive correlation mechanism between "activity level" and "resource supply" is established by dynamically mapping the token generation rate using linear affine transformation. This initialization method allows the frequency control system to move away from rigid static configurations and instead generate personalized flow control containers based on each user's real-time operating habits and historical load characteristics.

[0089] Further, obtaining the personalized frequency control upper limit at the current moment includes: statistically analyzing the actual token consumption of currently active virtual interaction vertices within a unit of time, calculating the ratio of the actual token consumption to the token generation rate, and determining it as the resource consumption saturation; querying the current available token count in the total family frequency control token pool, calculating the ratio of the current available token count to the capacity of the total family frequency control token pool, and determining it as the global remaining water level; dynamically configuring weight coefficients based on the network scheduling priority index in the advertising protocol parameters, using the weight coefficients to perform a weighted summation calculation of the resource consumption saturation and the complement of the global remaining water level, and obtaining the global load pressure coefficient; determining the negatively correlated mapping relationship curve according to the network scheduling priority index, mapping the global load pressure coefficient to a capacity scaling factor, multiplying the capacity scaling factor by the base bucket capacity to obtain the adjusted dynamic bucket depth, and determining the adjusted dynamic bucket depth as the personalized frequency control upper limit at the current moment. The specific process is as follows: Figure 3 As shown.

[0090] Specifically, the frequency control upper limit calculation process for the currently active virtual interaction vertex is initiated. First, at the end of the preset statistical period, the consumption counter of the token bucket object corresponding to the node is read to obtain the total amount of tokens actually consumed per unit time. The floating-point arithmetic unit is called to divide the total amount of tokens actually consumed by the token generation rate value set by the node during the initialization phase to calculate the resource consumption saturation that represents the current bandwidth occupancy of the node. At the same time, the global resource mutex lock in the shared memory area is accessed to read the current available token count value in the family total frequency control token pool. The current available token count value is divided by the preset total capacity value of the family total frequency control token pool to calculate the global remaining water level ratio. Then, the complement operation is performed, and the global remaining water level ratio is subtracted from the value 1 to calculate the global remaining water level complement that represents the overall resource scarcity of the family network.

[0091] Subsequently, a weighted calculation of the load pressure coefficient is performed. To reflect differentiated control, the network scheduling priority index of the currently active virtual interaction vertex is retrieved from the advertising protocol parameters. To achieve differentiated traffic control, this embodiment sets the network scheduling priority index to levels 1 to 10 (the smaller the value, the higher the priority), and divides it into three management tiers to configure weight coefficients:

[0092] High-priority tier (network scheduling priority index levels 1 to 3): It is determined that the sensitivity of this service to global pressure should be kept at a low level. The weight of the global remaining water level supplement is set to 0.3, and the weight of resource consumption saturation is set to 0.7.

[0093] Medium priority tier (network scheduling priority index levels 4 to 6): The weight of both is set to 0.5 to achieve an equal balance between local demand and global supply pressure.

[0094] Normal priority tier (network scheduling priority index greater than 6): Increase the weight of the global remaining water level complement to 0.8 and reduce the weight of resource consumption saturation to 0.2 to ensure that low-priority services are compressed first when overall resources are limited.

[0095] The weights mentioned above are multiplied by the corresponding saturation and water level complements, and then summed to generate the global load pressure coefficient.

[0096] Next, a negatively correlated adjustment mechanism is established by performing a nonlinear mapping of the capacity scaling factor: that is, as the global load pressure coefficient increases, the output capacity scaling factor decreases accordingly, the Sigmoid function instruction is called, and the adjustment parameters in the function are determined according to the network scheduling priority index. (Mapping center offset) and (Curve steepness):

[0097] ;

[0098] in, Indicates the capacity scaling factor. Indicates the global load stress coefficient. This indicates the preset mapping center offset. This indicates the preset curve steepness.

[0099] High-priority services (levels 1 to 3): The parameter is set to 0.8. Set to 6. Due to the offset High and steep The negative correlation curve is relatively low, and the decline is gradual. Even if the pressure coefficient is large, the calculated scaling factor can still remain above 0.8, protecting the core business from excessive contraction.

[0100] Medium priority services (levels 4 to 6): The parameter is set to 0.6. Set to 10.

[0101] Regular business (level 6 and above): The parameter is set to 0.4. When set to 16, the mapping curve exhibits high sensitivity. Once the pressure coefficient increases, the scaling factor will decrease sharply in a negative correlation, which can quickly narrow the upper limit to release resources.

[0102] Finally, the baseline bucket capacity of the node's dynamic token bucket is read, and the baseline bucket capacity is multiplied by the negatively correlated capacity scaling factor using a multiplier to obtain the adjusted dynamic bucket depth. Through an atomic write operation, the adjusted dynamic bucket depth value is written into the capacity control register of the token bucket object in real time, and it is confirmed as the personalized frequency control upper limit at the current moment. Through the above process, dynamic negative feedback adjustment of the token bucket capacity is realized, which actively shrinks as the system pressure increases and smoothly expands as the pressure decreases. Under the premise of ensuring the stability of system throughput, adaptive allocation of resources is achieved.

[0103] By comprehensively calculating the resource consumption saturation and the complement of the global remaining water level, a weighted summation algorithm is used to quantify the current overall pressure of the system. The pressure coefficient is nonlinearly mapped to a smooth capacity scaling factor, effectively avoiding numerical abrupt changes at critical points caused by linear adjustment methods. This mechanism enables the frequency control upper limit to smoothly converge or expand dynamically within the baseline capacity range according to the actual pressure, ensuring the overall throughput stability of the system while achieving fine-grained and adaptive scheduling of bandwidth resources.

[0104] Further, the step of checking whether there are available tokens in the dynamic token bucket and the family total frequency control token pool and performing operations includes: deploying a traffic interception hook at the application layer of the network protocol stack, using the traffic interception hook to capture the Hypertext Transfer Protocol request data packet for the advertising product, and suspending the request data packet to the kernel-mode blocking wait queue; reading the current token count register value of the dynamic token bucket and the global quota semaphore value of the family total frequency control token pool, and using an atomic instruction to simultaneously determine whether the current token count register value and the global quota semaphore value are both greater than 0; if the determination result is yes, then performing an atomic decrement operation to simultaneously deduct one unit of token count and global quota, releasing the request data packet from the blocking wait queue, and writing the request data packet into the send buffer of the network interface controller to perform physical transmission; if the determination result is no, then directly removing the request data packet from the blocking wait queue, constructing a rejection response message containing an excessive request status code, and returning the rejection response message to the initiator of the request data packet.

[0105] Specifically, a traffic interception hook function based on the Netfilter framework is deployed between the network layer and transport layer of the network protocol stack. When a data packet sent by the application layer is detected, the hook function parses the packet header information in real time. By matching the target port and protocol type, it identifies the Hypertext Transfer Protocol request data packet for the advertising product. In this embodiment, the sending process of the data packet is intercepted, its downward transmission in the protocol stack is suspended, and the socket buffer structure corresponding to the data packet is suspended to the kernel-mode blocking wait queue, forcing the sending process to enter a sleep waiting state. Immediately afterwards, the system enters the critical section to perform resource verification. The system accesses the shared memory area storing the dynamic token bucket state through a memory mapping mechanism, reads the current token count register value of the dynamic token bucket corresponding to the currently active virtual interaction vertex, and reads the global quota semaphore value of the family total frequency control token pool. It calls the processor's atomic comparison and swap instructions to simultaneously determine whether the current token count register value is greater than 0 and whether the global quota semaphore value is greater than 0 within the same clock cycle logic, thereby ensuring the consistency of state reading in a multi-threaded concurrent environment.

[0106] If the above atomicity checks all result in yes, an atomic decrement operation is immediately executed. Using an atomic subtraction instruction, the current token count register value of the dynamic token bucket is decremented by 1, and the global quota semaphore value of the home total frequency control token pool is decremented by 1. Subsequently, the corresponding process in the blocking waiting queue is woken up, the suspended request data packet is released, and it is restored to the transmission path of the protocol stack. The driver writes the data packet into the hardware transmission buffer of the network interface controller through the direct memory access mechanism, triggering the modulation and transmission of the physical signal.

[0107] If the above atomicity check fails, meaning either party's resources are exhausted, a rejection policy is executed. The memory resources occupied by the request packet are directly removed from the kernel's blocking wait queue, terminating its sending process. Simultaneously, a standard HTTP response message is constructed locally on the protocol stack, setting the status line to "HTTP / 1.1 429 Too Many Requests". This rejection response message is then injected directly back into the application layer's receive buffer. Upon receiving this message, the application layer parses the status code and terminates the retry logic for this advertising request.

[0108] To ensure the stability of the negative feedback regulation system, a smooth damping mechanism is built-in when performing capacity scaling. The Sigmoid mapping function is introduced, which makes the capacity scaling factor exhibit a smooth non-linear change near the load critical point, avoiding the numerical step caused by linear regulation at the threshold point. The register is updated in real time through atomic write operations, and a minimum adjustment cycle (such as 1 second) is set to prevent high-frequency jitter of the current limiting limit due to instantaneous traffic fluctuations. This ensures the throughput stability and system robustness of home network resources in complex concurrency scenarios.

[0109] By using atomic instructions to simultaneously judge and deduct local tokens and global semaphores, the risk of data races and resource overselling caused by inconsistent states during inspection and use in high-concurrency scenarios is completely eliminated. At the same time, by directly constructing rejection messages in kernel mode and blocking physical transmission, invalid traffic is avoided from occupying the transmission buffer of the network interface controller and wireless air interface resources. While ensuring strong consistency of frequency control strategy, the system's invalid transmission power consumption and network congestion probability are significantly reduced.

[0110] By constructing a heterogeneous graph of physical topology and behavioral logs and combining it with key biodynamic feature analysis, the system utilizes exponentially weighted moving average and throughput standard deviation to adaptively initialize parameters, and dynamically scales the frequency control upper limit based on global load pressure, achieving real-time matching of resource supply and user activity. Combined with kernel-level traffic interception and dual atomicity verification mechanisms, it effectively mitigates bandwidth competition in multi-device concurrent scenarios, significantly improving the utilization efficiency of home network resources and the overall throughput stability of the system while strictly protecting user privacy and business continuity.

[0111] Example 2:

[0112] When the system detects that a device is operating at high throughput, the background monitoring process calls the network interface controller driver to read the MAC address of the physical network card and performs a SHA-256 hash operation with a salt value to generate a unique hardware fingerprint. Simultaneously, the system samples the current wireless signal strength indicator value as -65dBm and combines it with the wireless access point service set identifier (WAS) of the wireless access point to construct network topology features. Concurrently, the input subsystem captures a continuous sequence of human-machine interface device commands. The system calculates the average physical interval between key presses in this sequence to be 120 milliseconds and extracts the process identifier of the current foreground application. The physical layer and application layer data are combined to generate a basic feature log.

[0113] The processor reads the aforementioned feature data and constructs a heterogeneous graph structure in memory, containing physical device vertices and virtual interaction vertices. The system calculates the deviation of the wireless signal strength index between two physical device vertices, maps this deviation to a signal-environment similarity weight of 0.92 using a Gaussian kernel function, and establishes an undirected connection. Subsequently, random walk sampling is performed based on semantic meta-paths, and nodes are mapped to 128-dimensional feature vectors using a skip-word model. The cosine similarity between nodes is calculated, and weighted isomorphic edges are established when the similarity value is higher than 0.7. A modularity maximization algorithm is run on the isomorphic graph. After multiple iterations, when the increment of the modularity Q-value converges to below 0.001, the clustered node set is locked into the same family unit.

[0114] Biodynamic analysis was performed on the behavioral data within the current time sliding window. The Shannon entropy of the key value distribution was calculated to be 1.8, and the jitter variance of the command response was 0.05 (indicating extremely stable operation). These features were concatenated with the one-hot encoded time slice vector and input into a gradient boosting decision tree classification model. The model performed multiple rounds of tree splitting calculations and output the original discrimination score belonging to the "youth user" category. The system performed Softmax normalization on this score, calculating the posterior probability value of this user's identity to be 0.94. Since this probability value exceeds the preset activation threshold of 0.85, the system immediately marked the corresponding virtual interaction vertex as active in the register.

[0115] For this active node, the system uses an exponentially weighted moving average algorithm (smoothing factor 0.2) to process the reciprocal of the instruction arrival interval, calculating the current interaction activity index. Simultaneously, it calculates the standard deviation of historical network throughput, multiplies it by a burst rate of 2.0, and calculates a baseline burst capacity of 960 tokens. Using a linear affine transformation function, the high activity index is mapped to a generation rate of 60 tokens per second. The system instantiates a token bucket object in memory, writes 960 to the bucket depth register, and writes 60 to the fill timer, completing the initialization.

[0116] During operation, the system monitors the load status in real time. The current resource consumption saturation of this node is statistically determined to be 0.85. Simultaneously, the system queries the overall family frequency control token pool and calculates the global remaining water level complement to be 0.9 (indicating extreme global resource strain). Using a preset weight (0.5, 0.5), the system calculates the global load pressure coefficient to be 0.875. Substituting this coefficient into the Sigmoid activation function, the system calculates the capacity scaling factor to be 0.70. A multiplication operation is then performed, multiplying the base capacity of 1200 by 0.70, resulting in an adjusted dynamic bucket depth of 672. The token bucket capacity limit is immediately updated to prevent single-point overload.

[0117] When the node initiates a new ad request packet, the kernel-mode traffic interception hook suspends it. The system executes atomic instructions while simultaneously reading the local token count and the global quota semaphore. If the local token count is greater than 0, but the global quota semaphore has been reduced to 0, the atomic check result is negative. The packet is then removed from the blocking queue, the socket buffer memory it occupied is released, and a rejection response message with status code 429 is constructed and returned to the application layer. No write operations are performed on the physical send buffer, thus completing precise flow control based on probabilistic identity and global pressure within milliseconds.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic advertising frequency control optimization method based on multidimensional assessment of family members, characterized in that, include: Obtain the network topology characteristics, account interaction logs, and advertising protocol parameters of shared devices, and obtain the total capacity of the family frequency control token pool; A device-node mapping graph is constructed based on network topology features and account interaction logs. The encrypted media access control address in the network topology features is mapped to physical device vertices, and the anonymized account unique identifier in the account interaction logs is mapped to virtual interaction vertices, ensuring the uniqueness of the graph vertices. Graph representation learning and community detection algorithms are used to determine family units. All physical device vertices and virtual interaction vertices in the same community module are identified as an independent family unit. Based on device input commands, the node mapping model is used to activate the corresponding currently active virtual interaction vertices from the family unit. The model outputs a probability value to represent the identity confidence of each virtual interaction vertex. The virtual interaction vertex with the highest probability value that exceeds a preset activation threshold is marked as active. If the probability values ​​are all below the activation threshold, the current state remains unchanged or it is marked as an unknown user. Obtain historical load data of currently active virtual interaction vertices, and calculate the interaction activity of nodes based on frequency and time span characteristics in account interaction logs; Initialize a dynamic token bucket for the currently active virtual interactive vertex; Based on historical load data and traffic burst tolerance in advertising protocol parameters, calculate the token bucket baseline capacity and calculate the token generation rate through interaction activity. The actual number of tokens consumed by currently active virtual interaction vertices per unit time is counted, and the ratio of the actual number of tokens consumed to the token generation rate is calculated and determined as the resource consumption rate. Query the current number of available tokens in the total family frequency control token pool, calculate the ratio of the current number of available tokens to the capacity of the total family frequency control token pool, and determine it as the global remaining water level; dynamically configure the weight coefficient based on the network scheduling priority index in the advertising protocol parameters, and use the weight coefficient to perform a weighted summation calculation on the resource consumption degree and the complement of the global remaining water level to obtain the global pressure coefficient; The negative correlation mapping curve is determined based on the network scheduling priority index. The global pressure coefficient is mapped to the capacity scaling factor. The capacity scaling factor is multiplied by the baseline bucket capacity to obtain the adjusted dynamic bucket depth. The adjusted dynamic bucket depth is determined as the personalized frequency control upper limit at the current moment. When a data distribution request for an advertising product is received, check whether there are available tokens in the dynamic token bucket and the family total frequency control token pool and perform the operation.

2. The dynamic advertising frequency control optimization method based on multidimensional evaluation of family members according to claim 1, characterized in that, The process of obtaining the network topology characteristics of shared devices, account interaction logs, and advertising protocol parameters of advertising products, and obtaining the total family frequency control token pool capacity includes: collecting the media access control address hash value, wireless access point service set identifier, and received signal strength indication of shared devices, and combining them to construct the network topology characteristics; capturing the sequence of human-machine interface device input instructions, application focus dwell time, and streaming media playback timestamps within a preset window to generate the account interaction logs; extracting the network scheduling priority index and traffic burst tolerance from the advertising product metadata, and determining them as the advertising protocol parameters; parsing the maximum number of requests per unit time and the token reset period in the global traffic quota configuration file, and calculating their product as the total family frequency control token pool capacity.

3. The dynamic advertising frequency control optimization method based on multidimensional evaluation of family members according to claim 1, characterized in that, The process of constructing the device-node mapping graph includes: traversing all physical device vertices; when the wireless access point service set identifiers of two physical device vertices are consistent, an undirected spatial edge is established between them; calculating the numerical deviation of the received signal strength indications of the two, and using a Gaussian kernel function to convert the numerical deviation into the signal environment similarity weight of the undirected spatial edge; traversing the account interaction logs; when a specific virtual interaction vertex is detected to generate a human-machine interface device command on a specific physical device vertex, a directed mapping edge is established between them, and the command trigger frequency per unit time is counted as the interaction strength weight of the directed mapping edge; combining the physical device vertices, virtual interaction vertices, undirected spatial edges, and directed mapping edges to generate a heterogeneous graph structure containing a node attribute matrix and a weighted adjacency matrix, which serves as the device-node mapping graph.

4. The dynamic advertising frequency control optimization method based on multidimensional evaluation of family members according to claim 1, characterized in that, The process of determining the family unit includes: defining a semantic meta-path from physical device vertex to virtual interaction vertex to physical device vertex in the device-node mapping graph, as a random walk path for constraint graph traversal; performing random walk sampling in the device-node mapping graph based on the semantic meta-path to generate a set of node sequences; training the set of node sequences using a skip word model with the goal of maximizing the co-occurrence probability of nodes within the sequence, mapping each vertex in the graph to a low-dimensional dense feature vector in a high-dimensional vector space; calculating the cosine similarity matrix between the low-dimensional dense feature vectors; constructing a weighted isomorphic graph based on the cosine similarity matrix; iteratively partitioning the weighted isomorphic graph using a modularity maximization algorithm until the modularity Q-value converges; and determining the set of vertices partitioned within the same module as the family unit.

5. The dynamic advertising frequency control optimization method based on multidimensional evaluation of family members according to claim 1, characterized in that, The activation of the corresponding currently active virtual interactive vertex includes: establishing a fixed-length time sliding window; serializing and sampling the device input command features within the sliding window; calculating the average time interval of key commands, the Shannon entropy of the key value probability distribution, and the jitter variance of the command response within the window; performing one-hot encoding on the current system time to generate time slice features; concatenating the average time interval, Shannon entropy, jitter variance, and time slice features into vectors to obtain a real-time operation feature vector; inputting the real-time operation feature vector into a pre-trained gradient boosting decision tree classification model to obtain the original discrimination score for each virtual interactive vertex; performing a normalized exponential function operation on the original discrimination score to calculate the posterior probability value; selecting the virtual interactive vertex with the highest posterior probability value that exceeds a preset activation threshold and marking its state as active.

6. The dynamic advertising frequency control optimization method based on multidimensional evaluation of family members according to claim 1, characterized in that, The process of initializing a dynamic token bucket for the currently active virtual interaction vertex includes: extracting the instruction timestamp sequence within the most recent preset period from the account interaction log, calculating the instruction arrival interval between adjacent instruction timestamps in the sequence, and taking the reciprocal of the instruction arrival interval as the instantaneous frequency feature; using an exponentially weighted moving average algorithm to iteratively smooth the instantaneous frequency feature to obtain a normalized interaction activity index; calculating the standard deviation of network throughput in the historical load data, multiplying the standard deviation, a preset burst multiplier coefficient, and the traffic burst tolerance in the advertising protocol parameters to determine the baseline burst capacity of the dynamic token bucket; using a linear affine transformation function to map the interaction activity index to the token generation rate of the dynamic token bucket; instantiating a new token bucket object, writing the baseline burst capacity into the bucket depth register of the current object, and writing the token generation rate into the token filling timer of the current object to complete the initialization of the dynamic token bucket.

7. The dynamic advertising frequency control optimization method based on multidimensional evaluation of family members according to claim 1, characterized in that, The process of checking whether there are available tokens in the dynamic token bucket and the family total frequency control token pool and performing operations includes: deploying a traffic interception hook at the application layer of the network protocol stack, using the traffic interception hook to capture the Hypertext Transfer Protocol request data packet for the advertising product, and suspending the request data packet to the kernel-mode blocking wait queue; reading the current token count register value of the dynamic token bucket and the global quota semaphore value of the family total frequency control token pool, and using an atomic instruction to simultaneously determine whether the current token count register value and the global quota semaphore value are both greater than 0; if the determination result is yes, then performing an atomic decrement operation to simultaneously deduct one unit of token count and global quota, releasing the request data packet from the blocking wait queue, and writing the request data packet into the send buffer of the network interface controller for physical transmission; if the determination result is no, then directly removing the request data packet from the blocking wait queue, constructing a rejection response message containing an excessive request status code, and returning the rejection response message to the initiator of the request data packet.

Images