An information collaborative screening method based on subscription matrix feature extraction

Abstract

The present application relates to big data processing technical field, specifically to a kind of information collaborative screening method based on subscription matrix feature extraction.Information multidimensional characteristic quantity is included and parsed into atlas node characteristic value;Fusion and concurrent load state quantity and subscription configuration instruction quantity, generate node fusion interaction quantity and node space coordination degree, extract feature matching degree parameter to construct subscription feature matrix;Through evaluation, obtain initial transmission potential energy value and comprehensive transmission potential energy value;Network node data stream greater than node transmission set threshold is gathered as target information cluster, and is stored in node cache data column;According to link interaction feedback quantity calculation strategy advantage residual scalar to update feature matching degree parameter.The present application is through the semantic atlas construction of information multidimensional characteristic quantity, based on node fusion interaction quantity Comprehensive transmission potential energy evaluation, and rely on link interaction feedback quantity Residual back propagation carries out depth integration, to meet the efficient collaborative screening of mass information.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, specifically to an information collaborative filtering method based on subscription matrix feature extraction. Background Technology

[0002] In today's large-scale distributed heterogeneous computing network environment, accurately extracting high-potential-value target feature clusters from massive, noisy, multimodal interactive data streams and achieving efficient business collaborative scheduling across computing nodes is a highly valuable technological direction. Existing technologies generally employ data filtering methods based on keyword indexing combined with hard-coded rule templates. This involves monitoring shallow character sequence changes in external heterogeneous data sources, and when a specific string or set tag is matched, automatically retrieving and extracting data packets that meet the set conditions from a pre-built central relational database or message queue. This is currently the mainstream technology for achieving large-scale network data extraction and distribution.

[0003] However, existing technical solutions have significant shortcomings in terms of semantic parsing depth and data distribution coordination. First, the index templates used for data filtering and matching suffer from shallowness and staticity. Existing indexing mechanisms are mostly based on fixed strings or historical business tags, and their matching boundaries cannot be adjusted in real time with the changes in the deep semantic environment of the external network data flow. This makes it difficult to accurately mine complex implicit data features that require deep logical judgment, resulting in a serious deviation in the range of target data cluster extraction. Second, the hard condition matching based on "Boolean logic" is only performed at the moment the query request is triggered. There is a lack of a quantitative evaluation mechanism for the affinity between the data flow and each distributed node, and it also fails to effectively consider the impact of the continuous fluctuation of the underlying network's concurrent load, resulting in rigid matching logic. The output of its data scheduling results suffers from siloing and redundancy. That is, it performs repeated independent retrieval and single-line distribution to all nodes in the computing network that issue similar requests, without combining the underlying topology logic for multi-path parallel adaptation scheduling. In addition, the existing technology severely severs the coupling relationship between data semantic parsing and underlying network concurrent scheduling. When facing high-load, multimodal heterogeneous network environments, it is difficult to meet the performance requirements for efficient collaborative filtering of massive amounts of information.

[0004] To address this, a collaborative information filtering method based on subscription matrix feature extraction is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an information collaborative filtering method based on subscription matrix feature extraction to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an information collaborative filtering method based on subscription matrix feature extraction, comprising: Data streams from network nodes are acquired through heterogeneous communication links, and multidimensional features are generated through multidimensional feature extraction. Semantic entity parsing and mapping relationship construction are performed on the multidimensional features to generate graph node feature values. The concurrent load state quantity and subscription configuration instruction quantity are obtained. The node fusion interaction quantity and node spatial synergy are generated by vector concatenation and inner product operation. Based on the node spatial synergy, the feature matching degree parameter is extracted from the feature value of the graph node, and the node fusion interaction quantity is introduced to perform cross-attention matching and dimension concatenation to generate the subscription feature matrix. A potential energy assessment is performed on the subscription feature matrix to generate an initial transmission potential energy value; the initial transmission potential energy value is combined with the node spatial synergy degree to perform topological weighting calculation and nonlinear activation operation to generate a comprehensive transmission potential energy value; The node transmission threshold is obtained and compared with the comprehensive transmission potential value. Network node data streams with comprehensive transmission potential values ​​exceeding the node transmission threshold are clustered to generate target information clusters, which are then synchronously loaded into the node cache data column. The link interaction feedback volume generated by the node cache data column is obtained, and a strategy advantage residual scalar is generated through residual backpropagation to update the feature matching degree parameters.

[0007] Preferably, the specific generation process of the information multidimensional feature quantity includes: performing data cleaning and word segmentation parsing operations on the network node data stream obtained from the heterogeneous communication link to generate discrete word sequence features; inputting the discrete word sequence features into a pre-trained deep bidirectional feature representation encoder, and performing multi-head self-attention feature mapping calculation to generate deep semantic embedding vectors; extracting concurrent timestamp identifiers and logical node topology identifiers carried in the network node data stream; performing time series embedding mapping on the concurrent timestamp identifiers to generate temporal fluctuation vectors; performing graph structure embedding mapping on the logical node topology identifiers to generate spatial location vectors; and performing tensor dimension alignment concatenation operation to fuse the deep semantic embedding vectors, temporal fluctuation vectors, and spatial location vectors to generate the information multidimensional feature quantity.

[0008] Preferably, the specific generation process of the graph node feature values ​​includes: inputting the multidimensional feature quantity of the information into a pre-configured named entity recognition computing engine, performing feature sequence decoding operations to generate underlying entity type data; combining the underlying entity type data to perform context dependency syntactic logic analysis to generate entity dependency association probability values; performing collision alignment calculations with a pre-set domain ontology knowledge base to extract node relationship edge weights; combining the underlying entity type data and node relationship edge weights to construct triplet topology data; inputting the triplet topology data into a graph attention neural network to perform neighbor node feature aggregation and multi-layer information transfer operations, and generating the graph node feature values ​​through nonlinear dimensionality reduction and smoothing calculations.

[0009] Preferably, the specific generation process of the node fusion interaction quantity and node spatial coordination degree includes: extracting the computing power allocation ratio value from the concurrent load state quantity and constructing a hardware computing power surplus vector; parsing the subscription configuration instruction quantity to extract the logical data request identifier and construct a task request dimension vector; inputting the hardware computing power surplus vector and the task request dimension vector into a preset aligned feedforward neural network to perform feature space projection mapping calculation, and outputting a computing power projection vector and a request projection vector with the same vector dimension; performing a low-level inner product dot product spatial operation on the computing power projection vector and the request projection vector to output a scalar calculation value; performing a normalization scaling operation on the scalar calculation value to generate the node spatial coordination degree; and fusing the computing power projection vector and the request projection vector to perform tail feature concatenation calculation to generate the node fusion interaction quantity.

[0010] Preferably, the specific generation process of the subscription feature matrix includes: obtaining feature matching degree parameters and splitting them into independent linear projection alignment weight matrices; using the linear projection alignment weight matrix to perform dimensional expansion transformation on the node fusion interaction quantity to generate a unified dimension attention query matrix; using the linear projection alignment weight matrix to perform hidden layer alignment transformation on the graph node feature values ​​to generate a unified dimension attention key value matrix and attention value matrix respectively; combining the attention query matrix and the attention key value matrix to perform scaling dot product space calculation to generate a cross-attention matching matrix; combining the cross-attention matching matrix with the attention value matrix to perform high-dimensional tiling expansion transformation to complete dimensional splicing and generate the subscription feature matrix.

[0011] Preferably, the specific generation process of the comprehensive transmission potential value includes: inputting the subscription feature matrix into a preset deep energy function model, performing a negative logarithmic derivation operation on the joint probability distribution of matrix elements to generate an initial transmission potential value in scalar form; extracting the logical routing level hop count of data communication scheduling; combining the logical routing level hop count with the node spatial coordination degree, performing a bottom-level multiplicative weighted attenuation topology calculation on the initial transmission potential value to generate a weighted superimposed potential value; inputting the weighted superimposed potential value into a hyperbolic tangent function calculation model, performing data boundary extreme value compression and state numerical regularization operations, and outputting the comprehensive transmission potential value within a fixed numerical threshold range.

[0012] Preferably, the specific generation process of the target information cluster includes: obtaining a node transmission set threshold, comparing it with a comprehensive transmission potential value, extracting network node data streams that exceed the node transmission set threshold and transferring them to a candidate scheduling processing queue; performing a low-level unsupervised clustering operation based on the feature density distribution law on the candidate scheduling processing queue to generate the target information cluster containing similar semantic logical structures; exploring distributed collaborative computing nodes with idle logical storage units in heterogeneous communication links; decomposing and dividing the target information cluster into fixed-length system data processing packets, and scheduling the data distribution control process to push the fixed-length system data processing packets concurrently, loading and storing them into the node cache data column divided inside the distributed collaborative computing node.

[0013] Preferably, the specific process for updating the feature matching degree parameter includes: collecting the physical cache hit ratio and computational resource waiting latency of the node cache data column during the data retrieval process, and fusing them to generate the link interaction feedback quantity; using the link interaction feedback quantity as the policy reward signal of the reinforcement learning framework, comparing the policy reward signal with the preset baseline value evaluation function, and performing advantage function difference deduction calculation to generate the policy advantage residual scalar; using the policy advantage residual scalar as the compensation weight coefficient of the loss function, and performing policy gradient chain derivation backpropagation along the feature value operation logic chain of the graph node to generate the underlying backpropagation gradient quantity; and using the underlying backpropagation gradient quantity in conjunction with the optimizer to perform iterative compensation update of the feature matching degree parameter.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By extracting multidimensional features to generate multidimensional information features, and then performing semantic entity parsing and mapping relationship construction, graph node feature values ​​are generated. This integrates semantic processing with graph structure features at the underlying level, constructing a multidimensional information representation system. In multimodal heterogeneous network environments, this system can deeply uncover the hidden logical connections within data flows and effectively establish topological mapping paths between data objects. This graph-based parsing method can accurately capture the core elements of high-dimensional data in the initial stage of information filtering, providing feature support for subsequent network interactions. This effectively bridges the coupling gap between high-dimensional data semantic parsing and underlying network concurrent scheduling, alleviating to some extent the deficiency of insufficient data distribution accuracy, and enabling a smooth and orderly handling of collaborative filtering tasks involving massive amounts of information.

[0015] 2. The system acquires concurrent load state data and subscription configuration command data, generates parameters such as node spatial coordination degree through vector concatenation and inner product operations, and then generates a subscription feature matrix. Potential energy assessment and topology weighted calculation are then performed to generate a comprehensive transmission potential energy value. This effectively integrates the real-time load status of the underlying physical hardware with high-level business requirements, constructing a decision framework for computing power awareness and data routing coordination. In high-load network environments, the potential energy assessment calculation enables the rational planning of data flow transmission paths and cache distribution, guiding information flow to suitable nodes. This helps alleviate the physical defects of low I / O addressing efficiency and excessive consumption of computing resources, promoting balanced resource allocation in massive data concurrency scenarios and steadily meeting the basic requirements for efficient collaborative screening and underlying physical performance scheduling.

[0016] 3. By acquiring the link interaction feedback volume generated by the node cache data column, and generating a policy advantage residual scalar through residual backpropagation, the feature matching degree parameters are updated, thus constructing a parameter adjustment closed-loop mechanism based on physical communication feedback. In a heterogeneous network environment, the physical interaction feedback of the underlying computing resources is captured and transformed into numerical basis to guide the iteration of the upper-layer feature network. By linking the parameter updates of high-dimensional data semantic parsing with the actual physical concurrent scheduling performance of the underlying network, a flexible and adaptive adjustment framework is established. During operation, it can gradually converge the logical deviations in the data distribution mechanism, continuously optimize the global execution state of the system, and enable the device to better adapt to the long-term underlying physical performance requirements of efficient collaborative screening of massive information, ensuring the stable operation of the overall execution logic.

[0017] 4. By deeply integrating the construction of a semantic graph of multidimensional information features, the comprehensive transmission potential assessment based on node fusion interaction, and the residual backpropagation relying on link interaction feedback, a global collaborative screening architecture is constructed. Through generating a subscription feature matrix and combining it with the strategy advantage residual scalar to perform closed-loop updates on feature matching parameters, dynamic alignment between the high-dimensional business semantic space and the underlying hardware logic space is achieved. This strongly coupled holistic architecture design effectively repairs the severe disconnect between high-dimensional data semantic parsing and underlying network concurrent scheduling. In the face of high-load, multimodal heterogeneous network environments, it effectively promotes the formation of a physical-level closed loop between macro-information recognition and micro-I / O addressing, standardizing the global execution logic. This helps to steadily resolve the global defects of excessive computing power consumption and insufficient data distribution accuracy, and can effectively meet the underlying physical performance requirements of modern computer equipment for efficient collaborative screening of massive amounts of information. Attached Figure Description

[0018] Figure 1 This is a flowchart of an information collaborative filtering method based on subscription matrix feature extraction proposed in an embodiment of this invention application; Figure 2This is a micro-flowchart of the multi-dimensional feature extraction and graph node generation proposed in an embodiment of the present invention. Figure 3 This is a flowchart of the feature matching degree parameter update method proposed in an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] Please see Figures 1-3 The present invention provides an information collaborative filtering method based on subscription matrix feature extraction, the specific steps of which are as follows: Data streams from network nodes are acquired through heterogeneous communication links, and multidimensional features are generated through multidimensional feature extraction. Semantic entity parsing and mapping relationship construction are performed on the multidimensional features to generate graph node feature values. The concurrent load state quantity and subscription configuration instruction quantity are obtained. The node fusion interaction quantity and node spatial synergy are generated by vector concatenation and inner product operation. Based on the node spatial synergy, the feature matching degree parameter is extracted from the feature value of the graph node, and the node fusion interaction quantity is introduced to perform cross-attention matching and dimension concatenation to generate the subscription feature matrix. A potential energy assessment is performed on the subscription feature matrix to generate an initial transmission potential energy value; the initial transmission potential energy value is combined with the node spatial synergy degree to perform topological weighting calculation and nonlinear activation operation to generate a comprehensive transmission potential energy value; The node transmission threshold is obtained and compared with the comprehensive transmission potential value. Network node data streams with comprehensive transmission potential values ​​exceeding the node transmission threshold are clustered to generate target information clusters, which are then synchronously loaded into the node cache data column. The link interaction feedback volume generated by the node cache data column is obtained, and a strategy advantage residual scalar is generated through residual backpropagation to update the feature matching degree parameters.

[0021] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0022] Example 1 This application discloses an information collaborative filtering method based on subscription matrix feature extraction, see below. Figure 1The specific steps proposed in this invention include: S1, acquiring network node data streams through heterogeneous communication links, and generating multi-dimensional information feature quantities through multi-dimensional feature extraction; performing semantic entity parsing and mapping relationship construction on the multi-dimensional information feature quantities to generate graph node feature values; S2, acquiring concurrent load state quantities and subscription configuration instruction quantities, generating node fusion interaction quantities and node spatial synergy through vector concatenation and inner product operations; based on node spatial synergy, extracting feature matching degree parameters from graph node feature values, and introducing node fusion interaction quantities for cross-attention matching and dimension concatenation to generate subscription feature moments. S3. Perform potential energy evaluation on the subscription feature matrix to generate an initial transmission potential energy value; combine the initial transmission potential energy value with the node spatial synergy degree to perform topological weighting calculation and nonlinear activation operation to generate a comprehensive transmission potential energy value; S4. Obtain the node transmission set threshold and compare it with the comprehensive transmission potential energy value. For network nodes whose comprehensive transmission potential energy value exceeds the node transmission set threshold, cluster the data streams to generate target information clusters and load them synchronously into the node cache data column; obtain the link interaction feedback amount generated by the node cache data column, generate the strategy advantage residual scalar through residual backpropagation, and update the feature matching degree parameters.

[0023] Furthermore, network node data streams are acquired through heterogeneous communication links, and multidimensional information features are generated through multidimensional feature extraction; semantic entity parsing and mapping relationship construction are performed on the multidimensional information features to generate graph node feature values; corresponding to step S1 above; refer to Figure 2 The specific implementation process includes: Data cleaning and word segmentation parsing operations are performed on the network node data stream obtained from the heterogeneous communication link to generate discrete word sequence features; the discrete word sequence features are input into a pre-trained deep bidirectional feature representation encoder, and multi-head self-attention feature mapping calculation is performed to generate deep semantic embedding vectors; concurrent timestamp identifiers and logical node topology identifiers are extracted from the network node data stream; time series embedding mapping is performed on the concurrent timestamp identifiers to generate temporal fluctuation vectors; graph structure embedding mapping is performed on the logical node topology identifiers to generate spatial location vectors; the deep semantic embedding vectors, temporal fluctuation vectors and spatial location vectors are fused and tensor dimension alignment concatenation operation is performed to generate the multidimensional feature quantity of the information.

[0024] Specifically, the generation process of multidimensional information features is as follows: Edge computing nodes receive data streams from network nodes via heterogeneous communication links. Since the data source contains unstructured communication text, semi-structured collaborative instructions, and serialized operational environment data, a multi-threaded data cleaning process is first instantiated. This process includes removing invalid tags, filtering out garbled segments, and stripping redundant header fields from the communication protocol. Subsequently, a byte-pair encoding-based word segmentation and parsing operation is employed. This operation segments the continuous unstructured data into sub-word structures by analyzing the probability of byte combinations in the data stream and maps them to a pre-defined underlying hash vocabulary space, thus transforming the original heterogeneous data stream into discrete word sequence features. These features serve as standard input data for the deep network model.

[0025] The generated discrete word sequence features are then fed into a pre-trained deep bidirectional feature representation encoder. This embodiment employs a transformer-based model. The encoder consists of multiple stacked encoding layers, each containing a multi-head self-attention computation module and a feedforward neural network module, with layers interconnected via residual connections and layer normalization.

[0026] To ensure the deep bidirectional feature representation encoder possesses the capability to extract features from heterogeneous office communication data, offline pre-training is required before integrating it into the online collaborative screening system. The pre-training dataset originates from heterogeneous communication log data deployed within the enterprise, including anonymized historical meeting communication texts, collaborative instruction records, and a mixed corpus of network environment data. Pre-training employs a masked language model as the objective function, randomly masking 15% of the words in the input discrete word sequence and training the model to predict the masked words by minimizing the cross-entropy loss function. The encoder uses a 12-layer Transformer encoder stacked structure, with 12 attention heads per layer, and the feedforward network's intermediate layers have a dimension of 3072. The pre-training process uses the AdamW optimizer, with an initial learning rate set to 1×10⁻⁶. -4 The batch size was set to 256, and the total number of training rounds was 30. After pre-training, the weights for extracting the bottom-level features were frozen, and online inference was performed only on the top-level output features.

[0027] In the multi-head self-attention feature mapping computation, discrete vocabulary sequence features are first projected through a linear transformation to generate a query matrix, a key matrix, and a value matrix. The multi-head mechanism computes the attention distribution in different subspaces in parallel, and the model extracts the distance dependencies between elements within the sequence through scaling dot product operations. Subsequently, a normalized exponential function is applied to the dot product results in the row dimension to calculate the probability weight distribution. Finally, this probability weight distribution is multiplied with the value matrix to generate and output a deep semantic embedding vector.

[0028] The specific dimension parameter of the deep semantic embedding vector is determined by cross-validation experiments on edge device memory capacity and processing latency. In the parameter configuration of this embodiment, the dimension parameter is set to 768 dimensions to represent the deep semantic information of the data stream in the hidden layer space.

[0029] The underlying message parser extracts the concurrent timestamp identifier and logical node topology identifier carried in the network node data stream.

[0030] For concurrent timestamp identifiers, a time series embedding mapping operation is performed. The specific calculation process for the time series mapping is as follows: for even-numbered dimensions in the vector, the result is obtained by dividing the value of the current concurrent timestamp by a specific power of a preset frequency period constant and then calculating its sine value, where the exponent of this power is obtained by dividing the current dimension index by the total number of dimensions in the time-domain hidden layer; for odd-numbered dimensions, the same independent variable is used to calculate the cosine value, thereby generating a time-domain fluctuation vector. The total dimension parameter of the hidden layer of the time-domain fluctuation vector is selected through ablation experiments according to the resolution granularity requirements of the time series, and in this embodiment, it is selected as 128 dimensions; the preset frequency period constant for controlling the frequency change of time wavelength is selected based on an empirical grid search for capturing multi-scale periods, and in this embodiment, it is set to 10000.

[0031] In this embodiment, before time-series mapping, the historical average concurrency density within the sliding time window of the current timestamp is extracted. This average density is injected as a phase offset factor into the trigonometric function variables of even and odd dimensions to construct a time-domain fluctuation vector containing local burst frequency information. Specifically, the length of the time sliding window is set; in this embodiment, the length is 500 milliseconds, and the total number of heterogeneous data stream packets arriving within the window is counted. The average local concurrency density is calculated by dividing the total number of heterogeneous data stream packets by the length of the time sliding window. An exponential smoothing filter is used to smooth the average density to obtain the current smoothed density value. Specifically, a smoothing coefficient of 0.3 is taken, the current average local concurrency density is multiplied by this smoothing coefficient, and then the product of the difference between 1 and the smoothing coefficient and the smoothed density value of the previous sliding window is added. The smoothed density value is divided by a preset maximum load-bearing density constant with the same dimensions to extract a dimensionless feature value. This constant is determined based on the historical stress test peak value of the office gateway, and in this embodiment, the constant is set to 20000. This dimensionless feature value is then normalized using the Sigmoid function to the interval between 0 and 1, generating a phase shift factor. In the original time series mapping calculation process, this phase shift factor is directly added to the independent variables of the sine and cosine functions. Specifically, for even-numbered dimensions, the calculation process is updated as follows: After extracting the timestamp (based on milliseconds) as a dimensionless pure number, it is moduloed by the system's natural day duration constant to obtain a relative time pure number reflecting intra-day periodic changes. This value is then divided by a specific power of 10000 (the exponent of which is the quotient obtained by directly dividing the current even-numbered dimension index by the total number of dimensions). The quotient obtained from the above division is added to the aforementioned phase offset factor as the target independent variable, and finally, the sine value of this target independent variable is calculated. For odd-numbered dimensions, the update involves calculating the cosine value of the same target independent variable. The final time-domain fluctuation vector is obtained through the above calculations. This process endows the time-domain features with the dynamic perception capability of burst traffic, effectively avoiding the homogenization of timestamp mapping under high concurrency.

[0032] For the logical node topology identifier, a graph structure embedding mapping operation is performed. An algorithm framework based on a random walk mechanism is adopted, sampling the local network structure neighborhood of the current data flow source node in the known network physical topology graph and training it to generate a spatial location vector. The dimension parameter of the spatial location vector is determined by performing principal component analysis on the number of topology nodes and testing the reconstruction error on a validation set. In this embodiment, the dimension of the spatial location vector is chosen to be 128 dimensions.

[0033] Finally, the deep semantic embedding vector, temporal fluctuation vector, and spatial location vector are fused, and a tensor dimension alignment concatenation operation is performed. The concatenation process involves connecting the three vectors along the last feature dimension of the tensor, and the final output is the multidimensional feature quantity of the information. The concatenated feature dimension is the sum of the three, i.e., 1024 dimensions.

[0034] By extracting deep semantics through an encoder and aligning and concatenating them with temporal fluctuations and spatial location vectors using tensors, the inherent context of the data stream and its physical spatiotemporal attributes are effectively integrated at the underlying level. This improves the completeness of multidimensional feature representation and provides structurally consistent data support for subsequent semantic parsing.

[0035] The multidimensional features of the information are input into a pre-configured named entity recognition computing engine to perform feature sequence decoding operations to generate underlying entity type data; context dependency syntactic logic analysis is performed in combination with the underlying entity type data to generate entity dependency association probability values; collision alignment calculation is performed between the entity dependency association probability values ​​and a pre-set domain ontology knowledge base to extract node relationship edge weights; triple topology data is constructed by combining the underlying entity type data and node relationship edge weights; the triple topology data is input into a graph attention neural network to perform neighbor node feature aggregation and multi-layer information transfer operations, and the graph node feature values ​​are generated through nonlinear dimensionality reduction and smoothing calculations.

[0036] Specifically, the generation process of the feature values ​​of the graph nodes is as follows: First, the multidimensional features of the information are input into a pre-configured named entity recognition computing engine. The underlying architecture of this engine uses a bidirectional long short-term memory network combined with a conditional random field. Feature sequence decoding operations are performed to generate underlying entity type data (such as recognizing entities like "project node" and "collaboration document").

[0037] Based on the extracted underlying entity type data, the dependency parsing component is then activated to perform contextual dependency parsing logic analysis. This module evaluates the logical relationships between different entities in the office collaboration instructions. The output of the analysis is an entity dependency association probability value, which is a continuous floating-point scalar used to quantify the likelihood of a specific logical dependency relationship between the identified entities.

[0038] The obtained entity dependency association probability values ​​are then used to perform collision alignment calculations with a pre-defined enterprise organizational structure and business collaboration ontology knowledge base. Specifically, this involves calculating the string distance between the text representation of the input entity pair and the standard entity template in the knowledge base. Based on the distance calculation result and the predefined association weights in the knowledge base, the entity dependency association probability values ​​are corrected. Specifically, the correction calculation method is as follows: the text similarity is obtained by subtracting the normalized string distance from 1; the text similarity, the predefined association weights, and the entity dependency association probability values ​​are then multiplied together, and the product result is extracted and output as the initial node relationship edge weights in the network. The aforementioned predefined association weights are assigned in a tiered manner based on the importance level of the business process; for example, the association weight for core business processes is assigned 1.2, and for ordinary auxiliary processes, it is assigned 0.8.

[0039] In this embodiment, the co-occurrence frequency of target entity pairs in the local cache within a preset historical time period is extracted, and dynamic semantic activity is calculated based on the co-occurrence frequency. This activity is used as a decay or gain coefficient to multiplicatively correct the extracted basic node relationship edge weights, and the updated node relationship edge weights are output. Specifically, a count-minimum sketch data structure is configured in the local node cache to record the co-occurrence frequency of each underlying entity pair within the most recent preset time period. In this embodiment, the preset time period is set to 24 hours. The baseline threshold for the co-occurrence frequency is set to 50 times. The dynamic semantic activity coefficient is calculated as follows: when the recorded co-occurrence frequency is greater than or equal to a baseline threshold, the quotient of the co-occurrence frequency divided by the baseline threshold is calculated, the logarithm of the quotient is taken to base 10, and 1 is added to this logarithm value as the dynamic semantic activity coefficient. When the co-occurrence frequency is less than the baseline threshold, the quotient of the co-occurrence frequency divided by the baseline threshold is calculated, 1 is subtracted from this quotient value to obtain the difference, and then the natural constant is raised to the power of this difference, which is taken as the dynamic semantic activity coefficient. The basic node relationship edge weights, previously calculated and corrected using the knowledge base, are obtained, and a multiplicative correction calculation is performed. Specifically, the basic node relationship edge weights, the aforementioned calculated dynamic semantic activity coefficient, and a penalty regularization term to prevent weight explosion are multiplied together. In this embodiment, the penalty regularization term is fixed at 0.95. If the corrected edge weight obtained from the above multiplication operation exceeds the set upper limit of 1.0, it is truncated to 1.0. Finally, this corrected edge weight is taken as the final output network node relationship edge weight. The above process introduces dynamic historical co-occurrence features, enabling the graph edge weights to adapt to the timeliness fluctuations of the underlying business flow.

[0040] By combining the underlying entity type data and the extracted node relationship edge weights, a triplet topological structure data is constructed. These triples are then transformed into a graph structure data model containing a set of nodes and a set of edges.

[0041] The constructed triplet topology data is input into a graph attention neural network model. During model processing, an attention mechanism is constructed to learn the attention coefficients between the target node and its neighboring nodes. Specifically, the calculation process involves concatenating the previous layer feature vectors of the target node and the previous layer feature vectors of its neighboring nodes, inputting this concatenation into a single-layer feedforward neural network, then passing it through a leaky rectified linear unit activation function to output a scalar. This scalar is then normalized using a normalized exponential function within the local neighborhood to generate the corresponding attention coefficients.

[0042] Specifically, the graph network feature aggregation and update calculation process for each hidden layer is as follows: First, extract the features of the previous layer of all neighboring nodes in the local neighborhood of the target node. Multiply these features by the corresponding attention coefficients and perform weighted summation. Then, multiply the original input features of the target node by a preset retention ratio parameter and add them to the above weighted summation result. Finally, input the overall summation result into the exponential linear unit activation function for nonlinear dimensionality reduction and smoothing calculation, thereby outputting the updated graph node feature value of the target node at the current level.

[0043] To clearly illustrate the parameter selection methods involved in the graph attention neural network model, the configuration is explained below: Network hiding layers: This parameter determines the range of node aggregation and is determined based on grid search verification to avoid feature homogenization within the cooperative gateway's memory constraints. This embodiment uses two layers.

[0044] Feature Preservation Ratio Parameter: This parameter controls the proportion of original input features in the current layer update. It is selected based on the best performance on the validation set to prevent oversmoothing; in this embodiment, it is 0.15.

[0045] Dropout rate: The proportion of neurons randomly dropped during training. This parameter is obtained by combining Bayesian optimization with network robustness cross-validation, and is 0.5 in this embodiment.

[0046] Before initial deployment, the aforementioned graph attention neural network underwent supervised training using a pre-defined labeled graph dataset. This dataset contains 5000 local triplet topology graphs from historical healthy office network segments. Node labels are categorized according to business entity type, including project nodes, collaborative documents, core meeting participants, and business instructions. Supervised training employs a multi-class cross-entropy loss function, with the classification of node entity types as the optimization objective. The training process utilizes the Adam optimizer with a learning rate of 0.005 and a weight decay parameter of 5 × 10⁻⁶. -4 The training cycle is 200 rounds. Through this supervised training, the model can adaptively amplify the weights of node features strongly related to the core business when aggregating neighbor features.

[0047] After the above iterative dimensionality reduction and smoothing calculations, the final graph node feature values ​​with a dimension of 256 are generated.

[0048] By constructing triples using entity recognition and syntactic analysis, and performing feature aggregation through graph neural networks, discrete semantics are transformed into graph feature values ​​with high-order topological associations, which enhances the ability to process implicit logical relationships and facilitates the capture of core mapping paths in complex heterogeneous data.

[0049] Further, the concurrent load state quantity and subscription configuration instruction quantity are obtained, and node fusion interaction quantity and node spatial synergy are generated through vector concatenation and inner product operation; based on the node spatial synergy, feature matching degree parameters are extracted from the graph node feature values, and the node fusion interaction quantity is introduced for cross-attention matching and dimension concatenation to generate the subscription feature matrix; corresponding to step S2 above; the specific implementation process includes: Extract the computing power allocation ratio from the concurrent load state variables and construct a hardware computing power reserve vector; parse the subscription configuration instruction quantity to extract the logical data request identifier and construct a task request dimension vector; input the hardware computing power reserve vector and the task request dimension vector into a preset aligned feedforward neural network to perform feature space projection mapping calculation, and output a computing power projection vector and a request projection vector with the same vector dimension; perform a low-level inner product dot product spatial operation on the computing power projection vector and the request projection vector to output a scalar calculation value; perform a normalization scaling operation on the scalar calculation value to generate the node spatial synergy degree; fuse the computing power projection vector and the request projection vector and perform tail feature concatenation calculation to generate the node fusion interaction quantity.

[0050] Specifically, the generation process of node fusion interaction volume and node spatial coordination degree is as follows: Extract concurrent load status metrics, including CPU idle ratio, computing power allocation ratio, and memory throughput. Normalize these metrics and concatenate them to construct a hardware computing power reserve vector. Simultaneously, parse subscription configuration commands (such as a command requesting "priority LAN synchronization of R&D department 3D drawing data"), extract logical data request identifiers, and map them to construct a task request dimension vector.

[0051] The hardware computing power margin vector and the task request dimension vector are respectively input into a pre-built aligned feedforward neural network. The feedforward network performs feature space projection mapping calculation and outputs computing power projection vectors and request projection vectors with the same vector dimension. Subsequently, L2 norm normalization is performed on the computing power projection vectors and request projection vectors respectively, and their magnitudes are uniformly constrained to 1. Then, the underlying inner product dot product space operation is performed on the normalized computing power projection vectors and request projection vectors, and the scalar calculation value (i.e., cosine similarity) constrained in the interval [-1, 1] is output.

[0052] The above scalar calculation values ​​are subjected to normalization scaling operations. The specific evaluation and deduction process is as follows: the transpose matrix of the computing power projection vector is extracted and multiplied by the requested projection vector to obtain a scalar representing the state similarity; then, the preset mean offset parameter is subtracted from the dot product scalar, and the difference is divided by the scaling temperature coefficient parameter used to control the slope of the curve. Finally, the negative number of the quotient is used as the exponent of the natural logarithm base for exponentiation, and one is added to it and the reciprocal of the whole is obtained to finally generate the node space coordination degree value constrained in a fixed interval.

[0053] Regarding parameter selection during the calculation process: The mean offset parameter is a mapping value obtained by proportionally equivalent conversion of the historical mean of the original unnormalized inner product space. It is used to forcibly separate data streams with different activity levels into the long-tail distribution range during the exponential scaling phase. In this embodiment, the mean offset parameter is set to 0.5.

[0054] Scaling temperature coefficient parameter: determined by comparing the accuracy of the judgment under different smooth slopes, which is 2.5 in this embodiment.

[0055] After generating the synergy degree, the computing power projection vector and the request projection vector are subjected to tail feature concatenation calculation, that is, a connection operation is performed along a specific dimension of the tensor. This generates a high-dimensional node fusion interaction quantity.

[0056] By employing a feedforward neural network to perform spatial projection on the hardware computing power margin and task request vector, the physical load and logical request with different dimensions are unified into the same dimension and then inner product is performed. This eliminates cross-domain computing barriers, gives the spatial synergy a clear mathematical meaning, and ensures the self-consistency of feature matching.

[0057] The feature matching degree parameters are obtained and segmented into independent linear projection alignment weight matrices. The linear projection alignment weight matrices are used to perform dimensional expansion transformation on the node fusion interaction quantities to generate a unified-dimensional attention query matrix. The linear projection alignment weight matrices are used to perform hidden-layer alignment transformation on the feature values ​​of the graph nodes to generate unified-dimensional attention key-value matrices and attention value matrices respectively. The attention query matrix and attention key-value matrix are combined to perform scaled dot product space calculation to generate a cross-attention matching matrix. The cross-attention matching matrix is ​​then combined with the attention value matrix to perform high-dimensional tiling expansion transformation to complete dimensional concatenation, generating the subscription feature matrix.

[0058] Specifically, the process of generating the subscription feature matrix is ​​as follows: Obtain the feature matching degree parameters, segment them, and transform them into three independent linear projection alignment weight matrices: set as query projection weight matrix, key projection weight matrix, and numerical projection weight matrix, respectively.

[0059] Using the query projection weight matrix, a dimensionality expansion transformation operation is performed on the node fusion interaction values ​​to generate a unified-dimensional attention query matrix. In parallel, using the key projection weight matrix and the numerical projection weight matrix, a hidden-layer alignment transformation operation is performed on the feature values ​​of the graph nodes to generate a unified-dimensional attention key-value matrix and attention value matrix.

[0060] Combining the attention query matrix and the attention key matrix, a scaled dot product space calculation is performed. The specific operation process of the cross-attention matching matrix is ​​as follows: obtain the attention query matrix and the transposed attention key matrix and perform matrix multiplication; divide each numerical element in the product matrix by the positive square root of the hidden layer feature dimension parameter to scale the values; then perform normalized exponential activation function operation on the elements of the scaled matrix row by row to generate a probability distribution weight matrix; finally, perform matrix multiplication between the probability distribution weight matrix and the attention value matrix to generate the cross-attention matching matrix.

[0061] The unified hidden layer feature dimension parameter in this step is selected based on the rank distribution and orthogonality requirements of the projection matrix. In this embodiment, the alignment is 256 dimensions.

[0062] The cross-attention matching matrix is ​​then combined with the attention value matrix to perform a high-dimensional tiling expansion transformation to complete the dimension splicing, and finally generate the subscription feature matrix.

[0063] The system utilizes independent linear projection weight matrices to generate query, key-value, and numerical matrices of a unified dimension, and then performs cross-attention matching. It also standardizes the dot product criteria of computing power features and graph features, avoiding the risk of dimension mismatch in heterogeneous matrices and improving the numerical stability of the subscription feature matrix generation.

[0064] Further, a potential energy assessment is performed on the subscription feature matrix to generate an initial transmission potential energy value; the initial transmission potential energy value is combined with the node spatial synergy degree for topological weighting calculation and nonlinear activation operation to generate a comprehensive transmission potential energy value; this corresponds to step S3 above; the specific implementation process includes: The subscription feature matrix is ​​input into a preset deep energy function model, and the joint probability distribution negative logarithmic derivation operation of the matrix elements is performed to generate an initial transmission potential value in scalar form. The logical routing level hop count of data communication scheduling is extracted. Combining the logical routing level hop count with the node spatial coordination degree, the initial transmission potential value is subjected to a bottom-level multiplicative weighted attenuation topology calculation to generate a weighted superimposed potential value. The weighted superimposed potential value is input into a hyperbolic tangent function calculation model, and data boundary extreme value compression and state numerical regularization operations are performed to output the comprehensive transmission potential value within a fixed numerical threshold range.

[0065] Specifically, the process for generating the comprehensive transport potential energy value is as follows: The subscription feature matrix is ​​input into a pre-set deep energy function model, which consists of a multi-layer fully connected network and a linear output layer. The model flattens the high-dimensional matrix and maps it to probability distribution parameters, then performs a negative logarithmic derivation of the joint probability distribution of the matrix elements to generate an initial transmission potential value in scalar form. Since this deep energy function model aims to minimize the energy output of healthy cooperative interaction states during offline training, the smaller the scalar value output by the model, the smaller the deviation between the current subscription feature matrix and historical high-quality cooperative states, indicating that the data stream has higher logical value and transmission priority. To adapt this energy value with inverse physical meaning to the forward scheduling threshold logic, the initial transmission potential value is inversely mapped in the subsequent hyperbolic tangent function calculation model by subtracting from a pre-set baseline bias parameter. This ensures that data streams with lower original energy values ​​have higher final mapped comprehensive transmission potential values.

[0066] In this embodiment, the information entropy of the row vectors of the subscription feature matrix is ​​extracted to construct a potential energy stability penalty term. This penalty term is used to smooth and calibrate the scalar form of the initial transmission potential energy value, filtering out potential energy spikes caused by local noise in the matrix, and outputting the calibrated initial transmission potential energy value. Specifically, each row vector of the 256-dimensional subscription feature matrix of the hidden layer is traversed. For the currently traversed row vector, the normalized exponential function probability distribution of each element is calculated, and then the information entropy of that row is calculated. The calculation method is as follows: extract the product of the probability value of each element and its natural logarithm, then sum the results of the products of all elements, and finally take the negative of the sum as the information entropy of that row. The average information entropy mean of the entire subscription feature matrix is ​​further calculated, that is, the information entropy of all row vectors is accumulated and divided by the total number of rows of the matrix. The standard information entropy reference value is set to 2.15. This reference value is obtained by statistically calculating the mathematical expectation of the information entropy of the feature matrix in a stable flow state in the offline logs of the enterprise gateway. The calculation logic yields a potential energy stability penalty term. The specific derivation process is as follows: Calculate the absolute value of the difference between the mean information entropy and the standard information entropy reference value. Divide this absolute value by the standard information entropy reference value to obtain the information entropy deviation rate. Subtract this information entropy deviation rate from 1 to obtain the penalty benchmark value. Then, take the larger of this penalty benchmark value and the hard lower limit value of 0.1 as the final potential energy stability penalty term. Obtain the initial transmission potential energy value originally generated by the deep energy function. Introduce the aforementioned penalty term to perform calibration calculations: Multiply the potential energy stability penalty term by a preset calibration tolerance coefficient (0.05 in this embodiment). Add 1 to the product result, and then multiply the result by the initial transmission potential energy value to obtain the final calibrated initial transmission potential energy value. Replace the original initial transmission potential energy value with this calibration result, and use it as the benchmark in the subsequent network topology weighting process. The above process reduces the interference of sudden changes in local feature noise on the potential energy assessment results.

[0067] The logical routing hop count for network communication scheduling is extracted. Specifically, the IP datagram headers in heterogeneous network data streams are parsed using a low-level packet sniffer to extract their Time-to-Live (TTL) field values. Combined with the default TTL constant initially set for the enterprise virtual private network, the physical route hop count from the source node to the current node is calculated using a difference operation, and this count is used as the logical routing hop count. Combining the enterprise-level router's logical routing hop count with the node spatial coordination degree, a low-level multiplicative weighted attenuation topology calculation is performed on the initial transmission potential value. Specifically, the weighted attenuation calculation process involves multiplying the initial transmission potential value by the node spatial coordination degree of the current device as the numerator, and simultaneously calculating a specific power of the natural logarithm constant as the denominator. The exponent of this specific power is set to the product of the logical routing hop count and a preset routing distance attenuation control parameter. Finally, the numerator is divided by the denominator to generate the weighted superimposed potential value.

[0068] The routing distance attenuation control parameter is derived from the statistical regression analysis of multi-hop delay in the enterprise communication network simulation model, and is set to 0.2 in this embodiment.

[0069] The weighted superimposed potential energy value is input into the hyperbolic tangent function calculation model, and data boundary extremum compression and state numerical normalization operations are performed. The above-mentioned state numerical normalization operation process is as follows: first, the calculated weighted superimposed potential energy value is multiplied by a preset gain adjustment multiplier, and a preset baseline bias term parameter is subtracted from the product; then, the hyperbolic tangent function value of the difference result is calculated; based on this, the hyperbolic tangent value is incremented by one; finally, the operation result is divided by two, thereby outputting the comprehensive transmission potential energy value within a fixed numerical threshold range.

[0070] The gain adjustment multiplier is selected based on experimental calibration of the sensitivity of the threshold determination region to potential energy changes; in this embodiment, it is chosen to be 1.2. The baseline bias term parameter is determined by combining the lower quartiles of the historical communication data distribution; in this embodiment, it is chosen to be 2.0.

[0071] By introducing a deep energy function model to perform negative logarithmic derivation and combining it with the number of routing hops for weighted attenuation calculation, a rigorous probabilistic and statistical basis for transmission potential energy is provided, which gets rid of the limitations of static threshold and enables the data flow transmission decision to have the flexible adjustment capability to dynamically adapt to the physical topology.

[0072] Further, a node transmission threshold is obtained and compared with the comprehensive transmission potential value. Network node data streams whose comprehensive transmission potential value exceeds the node transmission threshold are clustered to generate target information clusters, which are then synchronously loaded into the node cache data column. The link interaction feedback quantity generated by the node cache data column is obtained, and a strategy advantage residual scalar is generated through residual backpropagation to update the feature matching degree parameter; this corresponds to step S4 above; see reference. Figure 3 The specific implementation process includes: A node transmission threshold is obtained and compared with a comprehensive transmission potential value. Data streams from network nodes exceeding the node transmission threshold are extracted and transferred to a candidate scheduling processing queue. Unsupervised clustering based on feature density distribution is performed on the candidate scheduling processing queue to generate target information clusters containing similar semantic logical structures. Distributed collaborative computing nodes with idle logical storage units are searched in heterogeneous communication links. The target information clusters are decomposed and divided into fixed-length system data processing packets. The scheduling data distribution control process concurrently pushes these fixed-length system data processing packets and loads them into the node cache data columns partitioned within the distributed collaborative computing nodes.

[0073] Specifically, the generation process of the target information cluster is as follows: A node transmission threshold is obtained, which is dynamically selected based on the balance point between the false positive rate and the true positive rate using the receive operation characteristic curve. The overall transmission potential value is compared with the set transmission threshold. Network node data streams that are greater than or equal to the node transmission threshold are extracted and transferred to the candidate scheduling processing queue.

[0074] A low-level unsupervised clustering operation based on the feature density distribution pattern is performed on the candidate scheduling processing queue. This implementation uses a density-based spatial clustering algorithm to generate the target information clusters containing similar business logic structures.

[0075] Clustering algorithms rely on two parameters: the neighborhood radius and the minimum number of data points required. The parameters are selected as follows: Neighborhood radius parameter: By constructing a distance graph, the geometric distance from the collaborative data sample point in the candidate scheduling queue to the nearest neighbor is calculated, and the inflection point of the slope change on the curve is analyzed to determine the optimal radius threshold.

[0076] Minimum number of data points parameter: calculated based on the statistical expectation of the total number of candidate scheduling queue objects and the number of distributed points in the space of each neighboring object.

[0077] After clustering to generate target information clusters, the distributed collaborative computing nodes with idle logical storage units in the heterogeneous communication links are searched. The specific idle storage search criterion is as follows: the memory occupancy rate of the logical storage units of each collaborative computing node is periodically collected. When the memory occupancy rate is lower than a preset storage threshold, the node is determined to have idle logical storage units. In this embodiment, the preset storage threshold is set to 80%, suitable for receiving and loading fixed-length system data processing packets. The target information clusters are then decomposed and divided into fixed-length system data processing packets. The scheduling data distribution control process concurrently pushes the fixed-length system data processing packets and loads them into the node cache data columns partitioned within the distributed collaborative computing nodes.

[0078] Unsupervised clustering is performed on high-quality data based on feature density distribution, and fixed-length packets are pushed concurrently to idle computing nodes. This tightly integrates macroscopic semantic clustering with microscopic physical cache addressing, which reduces the probability of data congestion while improving the filling efficiency of node cache and the utilization rate of computing resources.

[0079] The physical cache hit ratio and computational resource waiting latency of the node cache data column during the data retrieval process are collected and fused to generate the link interaction feedback quantity. The link interaction feedback quantity is used as the policy reward signal of the reinforcement learning framework. The policy reward signal is compared with the preset baseline value evaluation function, and the advantage function difference inference calculation is performed to generate the policy advantage residual scalar. The policy advantage residual scalar is used as the compensation weight coefficient of the loss function. The policy gradient chain derivative is performed along the feature value operation logic chain of the graph node to generate the underlying backpropagation gradient quantity. The underlying backpropagation gradient quantity is used in conjunction with the optimizer to perform iterative compensation update of the feature matching degree parameter.

[0080] Specifically, the process of updating the feature matching degree parameter is as follows: The cache hit ratio and computational resource latency of the node cached data column during collaborative data retrieval are collected. These metrics are then fused to generate the link interaction feedback quantity. The calculation operation for generating the link interaction feedback quantity is as follows: the collected physical cache hit ratio value (a floating-point value constrained to the range of 0 to 1) is multiplied by a preset hit reward weight parameter to obtain the reward component; simultaneously, the collected latency value (based on milliseconds) is extracted as a dimensionless pure number, one is added, the logarithm to base 10 is calculated, and this logarithm is multiplied by a preset latency penalty factor parameter to obtain the penalty component; finally, the reward component is subtracted from the penalty component to obtain the link interaction feedback quantity.

[0081] The hit reward weight parameter and the delay penalty factor parameter are determined using the Pareto front multi-objective optimization algorithm in the reinforcement learning warm-up stage. In this embodiment, the hit reward weight parameter is set to 10.0 and the delay penalty factor parameter is set to 1.5.

[0082] By comparing the link interaction feedback quantity with the preset baseline value assessment function, a difference extrapolation calculation is performed to generate a strategy advantage residual scalar. The specific process of extrapolating and calculating the strategy advantage residual scalar involves obtaining the generated link interaction feedback quantity as the actual state return, and subtracting the baseline state assessment value output by the prior estimation of the baseline assessment model from it. This baseline assessment model is calculated based on historical link interaction feedback quantities using an exponentially weighted moving average, with a smoothing attenuation coefficient set to 0.95. The difference between the two is the strategy advantage residual scalar, which reflects the degree of advantage of the current routing scheduling strategy compared to the historical average level.

[0083] After generating the policy advantage residual scalar, the policy gradient chain-based derivation is performed along the feature value operation logic chain of the graph nodes to generate the underlying backpropagation gradient. The chain-based derivation calculation process of this underlying backpropagation gradient is as follows: First, the natural logarithm of the probability distribution of the policy decision under a specific office environment is calculated, and the partial derivative matrix of this logarithm with respect to the optimized feature matching degree parameter is solved. Then, each element in the obtained partial derivative matrix is ​​multiplied by the policy advantage residual scalar weight obtained above. Finally, the statistical expectation value of the above multiplication result is obtained within the overall data training batch range, and this expectation value is used as the generated backpropagation gradient compensation amount.

[0084] Finally, the feature matching degree parameters are iteratively updated using the backpropagation gradient from the underlying layer in conjunction with the optimizer. The core parameter configuration selection methods for the optimizer include: Initial learning rate: Selected based on observations of the smoothness of the validation set loss.

[0085] First-order and second-order momentum decay rates: determined based on the stability requirements of historical gradient exponential moving average and the need to control extreme gradients.

[0086] Numerical stability constant: determined according to the safety lower limit calculated by the underlying hardware floating-point method.

[0087] Training batch size: Determined through gradient accumulation testing based on the micro-memory capacity limit of the collaborative gateway node.

[0088] To ensure the real-time performance and high concurrency throughput of online collaborative information screening, the iterative compensation update mechanism for the feature matching degree parameters is executed asynchronously in the system background as an independent process. Specifically, when the accumulated feedback from line interactions reaches a preset batch threshold, such as 1000 samples, the system triggers an offline asynchronous derivative calculation and parameter update. The online data distribution routing process continues to use the currently fixed version of the feature matching degree parameters. After the background asynchronous update process completes the current batch iteration, the latest weight parameter matrix is ​​seamlessly loaded using memory pointer hot-switching technology. This decoupling design is to avoid delays and blockages caused by backpropagation calculations in online high-frequency data stream matching.

[0089] The physical cache hit rate and computation latency are transformed into reinforcement learning reward signals, and the gradient of the execution policy is back-calculated through the advantage function. The underlying discrete physical feedback is effectively connected to the upper-level continuous differentiation framework, which can self-correct feature weights according to the real physical state.

[0090] This embodiment elaborates on the underlying tensor architecture, dataset construction principles, offline training process, and core hyperparameters of the "aligned feedforward neural network" and "deep energy function model".

[0091] The alignment feedforward neural network is essentially a dual-tower feature space mapping architecture. Its input layer receives a 128-dimensional hardware computing power margin vector and a 128-dimensional task request dimension vector, respectively. The single-tower network architecture consists of three fully connected hidden layers connected in series, with neuron node dimensions set to 512, 256, and 256 respectively. Gaussian error linear units are used as non-linear activation functions between each hidden layer, and random deactivation layers with a dropout rate of 0.1 are inserted between layers to prevent feature overfitting. The final output layer of the network outputs a computing power projection vector and a request projection vector, both with a unified feature dimension of 256.

[0092] For the deep energy function model, the input is a subscription feature matrix generated by a cross-attention mechanism. The feature column dimension of this matrix is ​​fixed at 256 dimensions, while the row dimension is dynamically determined by the sequence length. The first layer of the model uses a global average pooling layer to calculate the mean along the sequence length, forcibly reducing the dynamic high-dimensional matrix to a one-dimensional feature vector of length 256. Subsequently, this one-dimensional feature vector passes through two fully connected hidden layers with 128 and 64 neurons respectively. These hidden layers are activated using leaky rectified linear units, where the slope of the negative half-axis is fixed at 0.01. Finally, the network ends with a single-node linear output layer without an activation function, directly outputting a single-dimensional scalar value as the basis for energy estimation.

[0093] To enable the model to possess physical feature recognition capabilities during online collaborative screening, offline pre-training was performed before initial deployment. The training dataset was derived from anonymized historical communication logs and physical node monitoring records accumulated by the enterprise gateway platform over the past twelve months.

[0094] For alignment feedforward neural networks, a contrastive learning offline dataset is constructed. Through log mining, collaborative task records with "data synchronization latency of less than 50 milliseconds and no packet loss" are extracted, and their corresponding computing power environment parameters are paired with task request parameters and marked as positive sample pairs. Unrelated computing power and task parameters from different time periods or different physical nodes are randomly extracted and randomly shuffled and combined, and marked as negative sample pairs.

[0095] For deep energy function models, a high-value feature matrix of successful local area network (LAN) distribution under stable network load is extracted from historical logs. Abnormal feature data caused by network jitter are removed, and an unsupervised single-class offline dataset containing only healthy and stable state data is constructed. After cleaning, the dataset is strictly divided into training, validation, and test sets in a ratio of 80%, 10%, and 10% respectively to ensure an objective evaluation of the model's generalization ability.

[0096] In the offline training process of the aligned feedforward neural network, an optimization logic based on triplet edge loss is adopted. The specific derivation process is as follows: In the same training batch, the computing power projection vector is extracted as an anchor point. The first Euclidean distance between this anchor point and the positive sample request projection vector, and the second Euclidean distance between this anchor point and the negative sample request projection vector are calculated. The difference between the second Euclidean distance and the first Euclidean distance is calculated, and this difference is subtracted from a preset decision boundary constant (set to 1.0 in this embodiment). If the result is greater than zero, it is used as the contrast loss value for the current batch; if it is less than or equal to zero, the loss value is zero. This loss value is minimized through the backpropagation algorithm, forcing the network to narrow the spatial angle between the matching physical computing power and the business request in the 256-dimensional hidden layer space, and widening the spatial distance of mismatched features.

[0097] In the offline training process of the deep energy function model, a hybrid loss derivation process based on reconstruction and negative log-likelihood is adopted. Since the model relies only on positive samples for unsupervised benchmark construction, its training objective is set to make the high-frequency healthy collaborative interaction matrix generate extremely low energy values. The specific loss function calculation process is as follows: calculate the natural exponent of the scalar energy value output by the model, and then take the negative logarithm of the exponent value as the basic loss; at the same time, an L2 regularization penalty term is introduced, that is, the mean of the sum of squares of the weight parameters of all neurons in the model is calculated; the basic loss is added to (the product of the weight decay coefficient and the regularization penalty term) and used as the global optimization loss for backpropagation, thereby forcing the model to learn the potential probability distribution boundary of the matrix features under normal business logic.

[0098] To ensure that the aforementioned dataset is fully utilized, a rigorous evaluation loop was constructed during the offline training process, with the core hyperparameters configured as follows: The first stage involves iterating based on the fundamental parameters of the training set. 80% of the training set data is extracted to perform multiple batches of forward inference and loss calculation. Since the model needs to find the statistical gradient direction among massive discrete samples, the batch size is uniformly set to 512. An adaptive moment estimation optimizer with a weight decay decoupling mechanism is used to execute the backpropagation algorithm to update the network weights. The initial learning rate of the aligned feedforward neural network is set to 5 × 10⁻⁶. -4 The initial learning rate for the deep energy function model is set to 1×10. -4 The weight decay factor is set to 1×10. -5 .

[0099] The second stage involves overfitting monitoring and dynamic parameter tuning based on the validation set. After each complete training epoch, the gradient updates of the network weights are temporarily frozen, and 10% of the validation set data is input into the current network to calculate the validation loss. This validation loss is an objective indicator that triggers dynamic parameter tuning: the learning rate decay strategy uses the cosine annealing algorithm, with a single annealing cycle set to 20 training epochs. An early stopping mechanism is also introduced, with a maximum of 150 training epochs for the entire pre-training process. If the validation set loss value does not show a significant decrease within 15 consecutive training epochs, the model is considered to have overfitting risk, and training is forcibly terminated. Simultaneously, the network weight matrix with the lowest historical validation loss is saved.

[0100] The third stage is the final blind test of generalization ability based on the test set. After the early stopping mechanism is triggered or the maximum number of training rounds is reached, the saved optimal weight matrix is ​​loaded, and a single independent blind test is performed on the test set, which accounts for 10% of the total and has not participated in training or parameter tuning. The objective generalization error of the model under completely unknown physical environment characteristics is calculated. Only when the global generalization error of the test set is lower than a safety threshold is the network weight matrix officially solidified and directly integrated into the online business collaborative screening architecture described above as the final "preset" model parameters.

[0101] This invention provides an information collaborative filtering method based on subscription matrix feature extraction. It constructs a cross-matching mechanism for computing-network convergence by deeply mapping the multidimensional semantic graph features of data streams with network concurrent load states. A potential energy assessment model is introduced during the data distribution phase to guide data flow to suitable computing node clusters. Simultaneously, based on the feedback from real physical cache interactions, residual backpropagation is used to perform closed-loop iteration of matching parameters. This globally linked architecture effectively repairs the disconnect between the high-dimensional business semantic space and the underlying hardware logic. In a multimodal heterogeneous network environment, it can reasonably avoid the risk of excessive consumption of computing resources and steadily improve overall I / O addressing efficiency, thereby effectively meeting the physical performance requirements for high-precision, adaptive collaborative filtering of massive amounts of information.

[0102] Example 2 This embodiment uses an information collaborative filtering method based on subscription matrix feature extraction applied to a distributed collaborative office system of a large, cross-regional enterprise as an example. In a wide-area remote office environment, employee devices (such as personal computing terminals and mobile smart devices) in different regions continuously generate massive heterogeneous network node data streams through virtual private networks and various wireless communication links. These streams include real-time video conferencing data, instant messaging text, high-frequency editing records of shared documents, and background environment synchronization commands. If all micro-operation records and multimedia streams are transmitted to the central main data center in real time without filtering, it will not only cause severe wide area network broadband congestion but also significantly increase the response latency of document collaborative editing and audio-visual interaction.

[0103] First, each office terminal, acting as a network edge access node, acquires data streams from the aforementioned heterogeneous office network nodes. Through data cleaning and word segmentation parsing, unstructured meeting chat logs and system logs are converted into discrete word sequence features. These are then input into a pre-trained deep bidirectional feature representation encoder for mapping calculation, generating deep semantic embedding vectors. Simultaneously, concurrent timestamps and logical topological locations (such as department network segments and project group identifiers) embedded in the terminal data packets are extracted, and temporal fluctuation vectors and spatial location vectors are independently generated through embedding mapping operations. Finally, these vectors are aligned and concatenated via tensor dimensions to generate multidimensional feature quantities reflecting the context of the current office behavior.

[0104] Next, semantic entity parsing is performed on the multidimensional feature to identify underlying entity types such as "project code," "report document," and "core attendees." Combined with context-dependent syntactic logic analysis, these related entities are collide-aligned with the enterprise's internal organizational structure and business process ontology knowledge base to extract node relationship edge weights and construct triplet topology data. This topology data is then input into a graph attention neural network to perform neighbor node feature aggregation and smooth dimensionality reduction calculations, generating graph node feature values ​​representing deep connections in business logic.

[0105] When assessing the data processing capabilities of local terminal devices, concurrent load status parameters such as the CPU idle rate and memory bus throughput of personal terminals are extracted. These parameters are then combined with subscription configuration commands issued by the collaborative office platform (e.g., "prioritize low-latency synchronization of 3D drawings from the R&D department"). Through feature space projection mapping and scalar multiplication calculations using a feedforward neural network, the node spatial collaboration degree representing the terminal's collaborative task processing status is deduced. Subsequently, computing power and task characteristics are integrated to generate node fusion interaction quantities with comprehensive perception capabilities.

[0106] The feature matching parameters are obtained and transformed into a projection alignment weight matrix. The node fusion interaction quantity and graph node feature values ​​are transformed into three sets of unified-dimensional attention query, attention key, and attention value matrices. By performing scaled dot product space calculation and multi-dimensional feature tiling and stitching, the macroscopic collaborative synchronization requirements are accurately associated with the underlying file features to generate a subscription feature matrix.

[0107] To determine whether a piece of collaborative data needs to be immediately transmitted across the network, the subscription feature matrix is ​​input into a deep energy function model, and a negative logarithmic derivation of the joint probability distribution is performed to generate an initial transmission potential value. Combining the expected logical routing hop count and node spatial coordination degree of the data packet in cross-border wide area network communication, a low-level multiplicative weighted attenuation calculation and data boundary extreme value compression operation are performed to generate a comprehensive transmission potential value within a fixed numerical range.

[0108] The system dynamically acquires and sets node transmission thresholds, comparing them with the overall transmission potential value. When the overall transmission potential value of a specific data stream (such as high-frequency document modification primitives for a key project team) exceeds the threshold, it is extracted and transferred to a candidate scheduling queue. For this queue, unsupervised clustering based on feature density is used to cluster massive amounts of scattered document modification records and communication fragments into highly structured target information clusters. Subsequently, the scheduling process searches for nearby colleague terminal devices with idle logical storage units within the same local area network, segments and packages the target information clusters, and prioritizes concurrently pushing them to the node cache data columns divided by office devices in the same group, realizing the local area network-based distribution and collaborative loading of high-value document instructions.

[0109] Finally, the underlying telemetry process collects in real time the physical cache hit rate of files during collaborative work and the computational resource waiting latency of document collaborative operations, and integrates them to generate the link interaction feedback. By comparing the residual scalar of the strategy advantage with the baseline value evaluation function, the strategy gradient chain derivative is performed along the operation logic chain to generate the gradient quantity in reverse and use the optimizer to iteratively compensate and update the feature matching degree parameter.

[0110] 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. An information collaborative filtering method based on subscription matrix feature extraction, characterized in that, include: Data streams from network nodes are obtained through heterogeneous communication links, and multidimensional feature quantities are generated through multidimensional feature extraction. Semantic entity parsing and mapping relationship construction are performed on multidimensional features of information to generate graph node feature values; The specific generation process of the multidimensional information feature quantity includes: performing data cleaning and word segmentation parsing operations on the network node data stream obtained from the heterogeneous communication link to generate discrete word sequence features; inputting the discrete word sequence features into a pre-trained deep bidirectional feature representation encoder, and performing multi-head self-attention feature mapping calculation to generate a deep semantic embedding vector; extracting the concurrent timestamp identifier and logical node topology identifier carried in the network node data stream; performing time series embedding mapping on the concurrent timestamp identifier to generate a temporal fluctuation vector; performing graph structure embedding mapping on the logical node topology identifier to generate a spatial location vector; fusing the deep semantic embedding vector, temporal fluctuation vector, and spatial location vector to perform tensor dimension alignment concatenation operation to generate the multidimensional information feature quantity; the logical node topology identifier includes the time-to-live field and network address field in the network data packet header; The concurrent load state and subscription configuration command quantities are obtained. Node fusion interaction quantities and node spatial synergy are generated through vector concatenation and inner product operations. Based on the node spatial synergy, feature matching parameters are extracted from the graph node feature values, and the node fusion interaction quantities are introduced for cross-attention matching and dimensional concatenation to generate the subscription feature matrix. The specific generation process of the subscription feature matrix includes: obtaining feature matching parameters and segmenting them into independent linear projection alignment weight matrices; using the linear projection alignment weight matrices to perform dimensional expansion transformation operations on the node fusion interaction quantities to generate a unified-dimensional attention query matrix; using the linear projection alignment weight matrices to perform hidden-layer alignment transformation operations on the graph node feature values ​​to generate a unified-dimensional attention key matrix and attention value matrix respectively; combining the attention query matrix and the attention key matrix to perform scaling dot product spatial calculation to generate a cross-attention matching matrix; combining the cross-attention matching matrix with the attention value matrix to perform high-dimensional tiling expansion transformation to complete dimensional concatenation, generating the subscription feature matrix. The subscription feature matrix is ​​subjected to potential energy evaluation to generate an initial transmission potential energy value. The initial transmission potential energy value is combined with the node spatial coordination degree to perform topology weighted calculation and nonlinear activation operation to generate a comprehensive transmission potential energy value. The specific generation process of the comprehensive transmission potential energy value includes: inputting the subscription feature matrix into a preset deep energy function model, performing negative logarithmic derivation operation on the joint probability distribution of matrix elements to generate an initial transmission potential energy value in scalar form; extracting the logical routing level hop count of data communication scheduling; combining the logical routing level hop count with the node spatial coordination degree, performing bottom-level multiplicative weighted attenuation topology calculation on the initial transmission potential energy value to generate a weighted superimposed potential energy value; inputting the weighted superimposed potential energy value into a hyperbolic tangent function calculation model, performing data boundary extreme value compression and state numerical regularization operation, and outputting the comprehensive transmission potential energy value within a fixed numerical threshold range. The node transmission threshold is obtained and compared with the comprehensive transmission potential value. Network node data streams with comprehensive transmission potential values ​​exceeding the node transmission threshold are clustered to generate target information clusters, which are then synchronously loaded into the node cache data column. The link interaction feedback volume generated by the node cache data column is obtained, and a strategy advantage residual scalar is generated through residual backpropagation to update the feature matching degree parameters.

2. The information collaborative filtering method based on subscription matrix feature extraction according to claim 1, characterized in that, The specific generation process of the graph node feature values ​​includes: inputting the multidimensional feature quantities of the information into a pre-configured named entity recognition computing engine, performing feature sequence decoding operations to generate underlying entity type data; combining the underlying entity type data to perform context dependency syntactic logic analysis to generate entity dependency association probability values; performing collision alignment calculations with a pre-set domain ontology knowledge base to extract node relationship edge weights; combining the underlying entity type data and node relationship edge weights to construct triplet topology data; inputting the triplet topology data into a graph attention neural network to perform neighbor node feature aggregation and multi-layer information transfer operations, and generating the graph node feature values ​​through nonlinear dimensionality reduction and smoothing calculations.

3. The information collaborative filtering method based on subscription matrix feature extraction according to claim 1, characterized in that, The specific generation process of the node fusion interaction quantity and node spatial coordination degree includes: extracting the computing power allocation ratio value from the concurrent load state quantity and constructing a hardware computing power surplus vector; parsing the subscription configuration instruction quantity to extract the logical data request identifier and construct a task request dimension vector; inputting the hardware computing power surplus vector and the task request dimension vector into a preset aligned feedforward neural network to perform feature space projection mapping calculation, and outputting a computing power projection vector and a request projection vector with the same vector dimension; performing a low-level inner product dot multiplication spatial operation on the computing power projection vector and the request projection vector to output a scalar calculation value; performing a normalization scaling operation on the scalar calculation value to generate the node spatial coordination degree; and performing tail feature concatenation calculation to generate the node fusion interaction quantity by fusing the computing power projection vector and the request projection vector.

4. The information collaborative filtering method based on subscription matrix feature extraction according to claim 1, characterized in that, The specific generation process of the target information cluster includes: obtaining a node transmission set threshold, comparing it with a comprehensive transmission potential value, extracting network node data streams that exceed the node transmission set threshold and transferring them to a candidate scheduling processing queue; performing a low-level unsupervised clustering operation based on the feature density distribution law on the candidate scheduling processing queue to generate the target information cluster containing similar semantic logical structures; exploring distributed collaborative computing nodes with idle logical storage units in heterogeneous communication links; decomposing and dividing the target information cluster into fixed-length system data processing packets, and scheduling the data distribution control process to push the fixed-length system data processing packets concurrently, loading and storing them into the node cache data column divided inside the distributed collaborative computing node.

5. The information collaborative filtering method based on subscription matrix feature extraction according to claim 1, characterized in that, The specific process of updating the feature matching degree parameter includes: collecting the physical cache hit ratio and computational resource waiting latency of the node cache data column during the data retrieval process, and fusing them to generate the link interaction feedback quantity; using the link interaction feedback quantity as the policy reward signal of the reinforcement learning framework, comparing the policy reward signal with the preset baseline value evaluation function, and performing advantage function difference inference calculation to generate the policy advantage residual scalar; using the policy advantage residual scalar as the compensation weight coefficient of the loss function, and performing policy gradient chain derivation backpropagation along the feature value operation logic chain of the graph node to generate the underlying backpropagation gradient quantity; and using the underlying backpropagation gradient quantity in conjunction with the optimizer to perform iterative compensation update of the feature matching degree parameter.

Images