Information matching management system based on big data

Abstract

The invention relates to the technical field of information processing, and discloses a big data-based information matching management system, which comprises an acquisition module, an attribute analysis module, a multi-dimensional modeling module and a dynamic matching module. The acquisition module acquires a user behavior data set and extracts active information demand features; the attribute analysis module analyzes the real-time updating frequency, the semantic association degree, the historical matching success rate and the data topological structure characteristics of the target information resources; a multi-dimensional modeling module performs heterogeneous modeling on each feature to generate timeliness, semantic association, matching probability feature vectors and topological structure vectors; and the dynamic matching module generates an adaptation degree score and determines an optimal matching object through heterogeneous space projection and feature coupling operation. The system improves the accuracy, real-time performance and efficiency of information matching through multi-dimensional feature processing, topological structure analysis and a dynamic matching algorithm, and is suitable for an accurate information matching scene in a big data environment.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to an information matching management system based on big data. Background Art

[0002] In the era of big data, the explosive growth of information has made the efficient matching of users and information resources a critical issue that needs to be addressed. Traditional information matching technologies often face multiple challenges and are unable to meet the precise matching requirements in complex scenarios.

[0003] From the perspective of information feature processing, existing systems are typically only able to perform simple processing on information features of a single or limited dimension, and are unable to effectively address the multi-dimensional, heterogeneous features of information resources, such as real-time update frequency, semantic relevance, historical matching success rate, and data topology. For example, for real-time update frequency, traditional methods may only use fixed-period statistical methods, ignoring the fluctuation patterns of data at different time scales, resulting in an inability to accurately capture the timeliness of information. In terms of semantic relevance processing, most systems are limited to keyword matching, lacking a deep understanding of contextual semantics and analysis of global thematic consistency. This results in superficial semantic matching and makes it difficult to explore the deeper semantic connections between information.

[0004] From the perspective of data topology analysis, traditional technologies lack effective modeling methods for the data topology of information resources. Data topology is a key reflection of the inherent connections within information resources. Its characteristics, such as core node density and edge weight distribution, can reveal the strength of connections and structural stability between information. However, existing systems often treat data as isolated entities and fail to fully consider their topological characteristics. This results in an inability to optimize the information matching process at the holistic level, making it difficult to handle information matching tasks within complex network structures.

[0005] In terms of feature fusion and matching algorithms, traditional matching algorithms often use simple linear combinations or fixed weights to fuse different features, lacking in-depth analysis of the dynamic relationships between features. For example, when processing time-sensitive features and topological structure features, the association weights between features cannot be dynamically adjusted based on real-time data, resulting in matching results that cannot accurately reflect the adaptability of information resources in different scenarios. Furthermore, when faced with multimodal and heterogeneous features, traditional algorithms lack effective cross-modal alignment mechanisms, making it difficult to achieve effective mapping and coupling between different feature spaces, thus affecting matching accuracy and efficiency.

[0006] With the continuous development of internet technology, users' information needs have become more dynamic and personalized. Traditional information matching systems have also shown significant deficiencies in real-time performance and dynamic adaptability. For example, when users' information needs change, traditional systems are unable to promptly adjust matching strategies based on their latest behavioral data, causing matching results to lag behind changes in user needs and impacting user experience. Furthermore, in the context of big data, the scale of information resources has expanded dramatically. Traditional systems' computational efficiency and scalability are insufficient to meet the demands of massive data processing, resulting in high computing resource consumption and slow matching speeds. Summary of the Invention

[0007] The purpose of the present invention is to provide an information matching management system based on big data to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an information matching management system based on big data, the system comprising: The acquisition module is used to obtain user behavior data sets and extract the information demand characteristics that are currently active; Attribute parsing module, used to parse the real-time update frequency, semantic relevance, historical matching success rate and data topology characteristics of target information resources; A multidimensional modeling module is used to perform heterogeneous feature modeling on the real-time update frequency, semantic relevance, historical matching success rate and data topology structure features, respectively, to generate a timeliness feature vector, a semantic relevance feature vector, a matching probability feature vector and a topology structure vector; The dynamic matching module is used to perform heterogeneous space projection and feature coupling operations on the timeliness feature vector, the semantic association feature vector, and the matching probability feature vector with the topological structure vector to generate a fitness score for each information resource and select the resource with the highest score as the optimal matching object.

[0009] Preferably, heterogeneous feature modeling is performed on the real-time update frequency, semantic relevance, historical matching success rate and the data topology structure features respectively to generate a timeliness feature vector, a semantic relevance feature vector, a matching probability feature vector and a topology structure vector, including: performing sliding period decomposition on the real-time update frequency to extract short-term fluctuation and long-term trend components and fusing them to generate the timeliness feature vector; performing context window analysis on the semantic relevance to extract local semantic similarity and global topic consistency and standardizing and splicing them to generate the semantic relevance feature vector; performing exponential smoothing prediction on the historical matching success rate to generate a probability change trajectory as the matching probability feature vector; and performing subgraph partitioning on the data topology to extract core node density and edge weight distribution as the topology structure vector.

[0010] Preferably, the data topology structure is divided into subgraphs to extract the core node density and edge weight distribution, including: using a community discovery algorithm to hierarchically divide the topology network, then counting the node centrality index and edge weight cumulative value of each subgraph, and generating a density-weight joint matrix as the topology structure vector.

[0011] Preferably, the timeliness feature vector, the semantic association feature vector and the matching probability feature vector are subjected to heterogeneous space projection and feature coupling operations with the topological structure vector to generate a fitness score for each information resource, including: performing cross-modal alignment based on a heterogeneous attention mechanism on the timeliness feature vector and the topological structure vector to generate a timeliness-topological association code; performing the cross-modal alignment based on a heterogeneous attention mechanism on the semantic association feature vector and the topological structure vector to generate a semantic-topological association code; performing the cross-modal alignment based on a heterogeneous attention mechanism on the matching probability feature vector and the topological structure vector to generate a probability-topological association code; calculating the timeliness contribution, semantic contribution and probability contribution respectively according to the timeliness-topological association code, the semantic-topological association code and the probability-topological association code; and generating the fitness score based on a linear combination of the timeliness contribution, the semantic contribution and the probability contribution.

[0012] Preferably, cross-modal alignment based on a heterogeneous attention mechanism is performed on the temporal feature vector and the topological structure vector to generate a temporal-topological association code, including: constructing a multi-granularity mapping tensor between the temporal feature vector and the topological structure vector, using an adaptive kernel function to perform feature transformation, and then generating the temporal-topological association code through a gated graph neural network.

[0013] Preferably, constructing a multi-granularity mapping tensor between the temporal feature vector and the topological structure vector includes: performing multi-resolution sampling on the temporal feature vector to generate layered feature slices, performing a tensor outer product operation on each layer of slices with the topological structure vector, and stacking them into the multi-granularity mapping tensor.

[0014] Preferably, an adaptive kernel function is used for feature transformation, including: dynamically selecting a bandwidth parameter of the kernel function according to the energy distribution of the multi-granularity mapping tensor, performing a nonlinear convolution operation, and then superimposing a jump connection to generate an optimized feature space.

[0015] Preferably, the temporality-topology association code is generated by a gated graph neural network, including: aggregating neighborhood node information from the optimized feature space input graph attention layer, and compressing it through a pooling operation after multi-layer propagation to generate the temporality-topology association code.

[0016] Preferably, the timeliness contribution, semantic contribution and probability contribution are calculated respectively according to the timeliness-topology association code, the semantic-topology association code and the probability-topology association code, including: inputting the timeliness-topology association code into a Gaussian process regression model, and calculating the timeliness contribution based on the prior covariance matrix; inputting the semantic-topology association code into a capsule network, and iteratively generating the semantic contribution through a dynamic routing mechanism; inputting the probability-topology association code into a random forest classifier, and outputting the probability contribution based on feature importance sorting.

[0017] Preferably, the output of the probability contribution based on the feature importance sorting includes: performing Shapley value decomposition on the probability-topology association code to calculate the contribution weight of each dimension, smoothing through exponentially weighted moving average, and finally using S-type function compression to generate a probability contribution in the range of 0 to 1.

[0018] Compared with the prior art, the present invention has the following beneficial effects: In terms of information feature processing, the acquisition module acquires user behavior datasets and extracts information demand characteristics of active states. The attribute parsing module comprehensively analyzes the target information resource's real-time update frequency, semantic relevance, historical matching success rate, and data topology. The multidimensional modeling module performs heterogeneous feature modeling for each of these features. For example, it performs a sliding-period decomposition of the real-time update frequency, extracts short-term fluctuation and long-term trend components, and fuses them to generate a timeliness feature vector. This processing method more accurately captures the temporal dimension of information and, compared to traditional fixed-period statistical methods, better reflects the variation patterns of data at different time scales, providing a solid foundation for accurately assessing information timeliness. Contextual window analysis of semantic relevance extracts local semantic similarity and global topic consistency, and standardized concatenates these to generate a semantic relevance feature vector. This breaks through the limitations of traditional keyword matching, deeply exploring the deep connections between contextual semantics and global topic consistency, making the expression of semantic features more comprehensive and accurate, and improving the accuracy of semantic matching.

[0019] In terms of data topology analysis, a community discovery algorithm is used to hierarchically segment the topological network. The node centrality index and edge weight accumulation values ​​of each subgraph are calculated to generate a joint density-weight matrix as the topological structure vector. This method fully considers characteristics such as the density of core nodes and the distribution of edge weights in the data topology, treating information resources as inherently interconnected network structures rather than isolated individuals. This method can reveal the strength of associations and structural stability between information at the overall structural level, providing a more comprehensive structural feature basis for information matching. This effectively addresses the problem of traditional technologies neglecting data topology and enables the matching process to better adapt to the information matching needs within complex network structures.

[0020] In terms of feature fusion and matching algorithms, the dynamic matching module uses a cross-modal alignment method based on a heterogeneous attention mechanism to process different feature vectors and topological structure vectors, generate association codes, and calculate the temporal contribution, semantic contribution, and probabilistic contribution respectively. Finally, a fitness score is generated through linear combination. For example, a multi-granularity mapping tensor is constructed and an adaptive kernel function is used for feature transformation, combined with a gated graph neural network to generate association codes. This approach can dynamically capture the multi-granularity association relationships between features, adaptively adjust the kernel function parameters according to the energy distribution of the data, optimize the feature space, and effectively improve the efficiency and accuracy of feature fusion. At the same time, different algorithms such as Gaussian process regression models, capsule networks, and random forest classifiers are used to calculate the contribution of different association codes. This fully utilizes the advantages of different algorithms in processing different types of data, enabling a more accurate assessment of the contribution of features in each dimension, thereby generating a more reasonable fitness score, ensuring that the highest-scoring resource is selected as the optimal match, and significantly improving the accuracy of information matching.

[0021] In terms of dynamic adaptability and real-time performance, the system can dynamically adjust feature modeling and matching strategies based on real-time user behavior data and updates to information resources. For example, exponential smoothing predictions are applied to historical matching success rates to generate probability change trajectories, which promptly reflect dynamic trends in matching success rates and enable the system to adjust subsequent matching strategies based on the latest matching results. Furthermore, the entire matching process utilizes efficient algorithms and models, enabling rapid processing of massive amounts of information in a big data environment. This ensures real-time matching results, meets users' dynamic and personalized information needs, and enhances the user experience.

[0022] Furthermore, by integrating and deeply processing multi-dimensional, heterogeneous features, the system effectively improves the efficiency and quality of information matching, reducing the time and effort users spend sifting through massive amounts of information for effective resources and providing users with a more convenient and efficient way to acquire information. Furthermore, the system is highly scalable, adapting to information matching scenarios across diverse fields and types, providing a wider range of technical support and solutions for information processing and applications in the big data era. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a working principle diagram of the information matching management system based on big data according to the present invention; Figure 2 Design graph generated for topological structure vector; Figure 3 Design diagram for heterogeneous space projection operation; Figure 4 Design graph generated for temporality-topology association coding; Figure 5 This is a design diagram of the contribution calculation process. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] See also Figure 1-Figure 5 The present invention relates to an information matching management system based on big data, and its specific implementation scheme is as follows: The system includes an acquisition module, an attribute analysis module, a multidimensional modeling module and a dynamic matching module.

[0026] The collection module acquires user behavior datasets and extracts the characteristics of currently active information needs. User behavior datasets contain user operation data such as clicks, browsing, searching, and favorites on various terminal devices, and are acquired in real time through tracking technology and log collection. Information need characteristics can be extracted through natural language processing techniques to perform keyword extraction and semantic analysis on user-entered query text, or through user behavior sequence analysis to infer current needs. For example, active demand types can be determined based on high-frequency operations within a time window.

[0027] The attribute parsing module analyzes the target information resource's real-time update frequency, semantic relevance, historical matching success rate, and data topology. Real-time update frequency is determined by monitoring the resource's version change timestamps, content increment frequency, and other statistical data. Semantic relevance is measured by calculating cosine similarity and co-occurrence relationships between the resource's content and a pre-defined thesaurus. Historical matching success rate is based on historical matching records between the resource and user needs, such as the ratio of successful matches to total matches. Data topology is extracted by constructing an entity-relationship graph of the resource, for example, by treating entities in the text as nodes and relationships between entities as edges to form a topological network.

[0028] The multidimensional modeling module performs heterogeneous feature modeling on real-time update frequency, semantic relevance, historical matching success rate and data topology structure features respectively, and generates timeliness feature vector, semantic relevance feature vector, matching probability feature vector and topology structure vector.

[0029] The dynamic matching module projects the temporal, semantic, and matching probability feature vectors into heterogeneous spaces and performs feature coupling operations on the topological structure vector. This generates a fitness score for each information resource and selects the resource with the highest score as the optimal match. Specifically, cross-modal projection technology is used to map vectors from different feature spaces into a unified semantic space. A feature coupling algorithm is then used to calculate the interactions between features in each dimension, ultimately generating a comprehensive score through weighted summation and other methods.

[0030] The present invention will be further described below in conjunction with Examples 1 to 5: Example 1: In the multidimensional modeling module, heterogeneous feature modeling is performed for real-time update frequency, semantic relevance, historical matching success rate, and data topology structure features to generate corresponding feature vectors.

[0031] Sliding-cycle decomposition is performed on the real-time update frequency to extract short-term fluctuation and long-term trend components, which are then integrated to generate a time-sensitive feature vector. This sliding-cycle decomposition employs the seasonal decomposition model used in time series analysis to decompose the real-time update frequency data into a trend term, a seasonal term, and a residual term. Short-term fluctuations correspond to the seasonal and residual terms, while long-term trends correspond to the trend term. After normalization, each component is concatenated into a feature vector along the time dimension. For example, with a daily period, the daily fluctuation pattern of the update frequency is extracted as a short-term feature, while with a monthly period, the trend change in the update frequency is extracted as a long-term feature. When performing sliding-cycle decomposition, the appropriate decomposition window size must be determined. This depends on the update characteristics of the information resource and the time sensitivity of user needs. For information resources with a high update frequency, such as news, a smaller window size can be used; for information resources with a low update frequency, such as academic papers, a larger window size can be used. During normalization, the trend term, seasonal term, and residual term are standardized to eliminate dimensional differences between the different components. The normalized components are arranged in chronological order to form a multidimensional vector, where each dimension corresponds to the decomposed component value at a specific point in time. To enhance the expressive power of the feature vector, timestamp information can be added to the vector to indicate the acquisition time of each data point. In this way, the time-sensitive feature vector can fully reflect the dynamic nature of information resource updates, providing strong support for subsequent matching operations.

[0032] Contextual window analysis is performed on semantic relevance to extract local semantic similarity and global topic consistency, which are then normalized and concatenated to generate a semantic relevance feature vector. Contextual window analysis uses a sliding window technique, setting a fixed-length window within the information resource text. The cosine similarity between the text within the window and the user's desired keywords is calculated as the local semantic similarity. Global topic consistency uses the latent Dirichlet allocation model to extract the topic distribution of the text. The KL divergence between each topic and the user's desired topic is calculated as the consistency metric. After normalizing the local similarity and global consistency metrics, the data are concatenated in the order of the windows to form a feature vector. Choosing the window size is crucial when implementing contextual window analysis. A window that is too small may not capture sufficient contextual information, resulting in inaccurate local semantic similarity calculations. A window that is too large may introduce excessive irrelevant information, affecting the quality of the feature vector. Generally speaking, the window size can be adjusted based on the length and semantic complexity of the text. For short texts such as microblogs and comments, the window size can be set to 10-20 words; for long texts such as articles and reports, the window size can be set to 50-100 words. When calculating local semantic similarity, the text within the window and the user demand keywords must first be preprocessed, including operations such as word segmentation, stop word removal, and word vectorization. Then, by calculating the cosine similarity between the word vectors, the semantic similarity score between the text within the window and the user demand keywords is obtained. For the calculation of global topic consistency, the latent Dirichlet allocation model is used to perform topic modeling on the information resource text to obtain the topic distribution of the text. At the same time, topic modeling is also performed on the user demand text to obtain the topic distribution of the user demand. By calculating the KL divergence between the two topic distributions, the consistency between the text and the user demand at the topic level is evaluated. After normalizing the local semantic similarity and global topic consistency indicators, they are spliced ​​into a vector according to the window order. To improve the representation ability of the feature vector, it is also possible to consider adding position weights and assigning different weights to windows at different positions to highlight the semantic information of important positions.

[0033] Exponential smoothing is performed on the historical matching success rates to generate a probability trajectory, which serves as the matching probability feature vector. Exponential smoothing uses triple exponential smoothing to smooth the historical matching success rate series, capturing its trend and seasonal variations. This method then predicts success rate values ​​for future time periods, forming a time series probability trajectory. When implementing exponential smoothing, the first step is to determine an appropriate smoothing parameter. The value of the smoothing parameter directly affects the sensitivity of the forecast results to historical data. A larger smoothing parameter results in forecasts that are closer to recent data and more responsive to changes, but may overlook long-term trends in the historical data. A smaller smoothing parameter results in smoother forecasts that are more dependent on historical data but may lag behind actual changes. Generally, the smoothing parameter can be optimized based on historical data to minimize forecast error. When performing triple exponential smoothing, the level, trend, and seasonal values ​​of the series are smoothed separately. These values ​​are iteratively updated to generate forecasts for future time periods. The predicted success rate values ​​are arranged in chronological order to form a time series vector. To enhance the expressive power of the vector, forecast confidence intervals can also be included to indicate the uncertainty of the forecast results. In this way, the matching probability feature vector can reflect the changing trend of the possibility of information resources matching user needs in the future time period.

[0034] The data topology is subgraphed to extract core node density and edge weight distribution as topological structure vectors. A community discovery algorithm is used to hierarchically segment the topological network into multiple subcommunities. For each subgraph, the mean of the node centrality index is calculated as the core node density, and the ratio of the cumulative sum of edge weights to the number of edges is calculated as the edge weight distribution index. The density and weight indexes of each subgraph are concatenated into a vector in the order of the subgraph numbers. When implementing subgraph partitioning, selecting an appropriate community discovery algorithm is crucial. Different community discovery algorithms are suitable for different types of network structures. For example, the Louvain algorithm is suitable for large-scale networks and has high computational efficiency, while the GN algorithm is suitable for discovering hierarchical community structures. The most suitable algorithm for subgraph partitioning is selected based on the characteristics of the information resource topological network. When calculating the core node density, multiple centrality indicators are used to comprehensively assess the importance of nodes, including degree centrality, betweenness centrality, and closeness centrality. The core node density of each subgraph is obtained by calculating the mean of these indicators. To calculate the edge weight distribution, we first count the weights of all edges within the subgraph. We then calculate the cumulative sum of these weights and divide it by the number of edges to obtain the average edge weight. The core node density and average edge weight of each subgraph are arranged in order of subgraph number to form a topological structure vector. To preserve the hierarchical relationship between subgraphs, we can add a hierarchical encoding to the vector to indicate the position of each subgraph in the hierarchical partitioning.

[0035] Through the above heterogeneous feature modeling of real-time update frequency, semantic relevance, historical matching success rate, and data topology, we generate a timeliness feature vector, a semantic relevance feature vector, a matching probability feature vector, and a topology structure vector. These feature vectors characterize the characteristics of information resources from different dimensions, providing a rich and comprehensive feature representation for subsequent dynamic matching operations. In practical applications, the feature modeling method can be adjusted and optimized based on specific business needs and data characteristics to improve the accuracy and efficiency of information matching. For example, for application scenarios with high timeliness requirements, the weight of the timeliness feature vector can be increased; for application scenarios with high semantic matching requirements, the optimization of the semantic relevance feature vector can be strengthened. By continuously improving and refining the feature modeling method, we can better meet the information needs of different users and enhance the performance and user experience of the information matching management system.

[0036] Example 2: When subgraphing the data topology to extract core node density and edge weight distribution, a community discovery algorithm is used to hierarchically partition the topological network. The degree centrality, betweenness centrality, and closeness centrality of each node are first calculated. Principal component analysis is then used to reduce these multi-dimensional centrality metrics to a single comprehensive centrality metric. The comprehensive centrality metrics of the nodes within each subgraph are then averaged to obtain the core node density of that subgraph. Edge weight distribution is calculated by summing the edge weights of all edges within the subgraph and dividing them by the number of edges to obtain the average edge weight. The core node density and average edge weight of each subgraph are arranged in the order of the hierarchical partitioning to generate a joint density-weight matrix. The rows of this matrix correspond to the subgraph number, and the columns correspond to the density and weight metrics. Vectorization is then used to convert this matrix into a one-dimensional topological structure vector.

[0037] When implementing subgraph partitioning, the first step is to construct a topological network of information resources. This network uses information entities as nodes and the relationships between entities as edges. The weight of the edge represents the strength of the relationship. For example, in an academic literature network, nodes can represent entities such as papers, authors, and institutions, while edges can represent citations between papers and collaborations between authors. Edge weights can be determined based on factors such as the number of citations and the frequency of collaboration. After constructing the topological network, an appropriate community discovery algorithm is selected for hierarchical segmentation. The community discovery algorithm used here can partition the network into multiple levels of subcommunities, where the connections between nodes within each subcommunity are relatively dense, while the connections between subcommunities are relatively sparse.

[0038] When calculating node centrality metrics, degree centrality measures the number of connections to a node, that is, the number of edges directly connected to the node. A higher degree centrality indicates a more active node in the network and a greater number of direct connections with other nodes. Betweenness centrality measures the importance of a node as a bridge in the network, that is, the number of shortest paths passing through the node. A higher betweenness centrality indicates a greater intermediary role for the node in information dissemination and resource circulation. Closeness centrality measures a node's proximity to other nodes in the network, that is, the reciprocal of the average shortest path length from the node to all other nodes. A higher closeness centrality indicates that the node can quickly exchange information with other nodes in the network.

[0039] After calculating each node's degree centrality, betweenness centrality, and closeness centrality, these metrics reflect the importance of the node in the network from different perspectives, but there may be some correlation between them. To eliminate this correlation and simultaneously integrate multi-dimensional information into a single metric, principal component analysis (PCA) is used for dimensionality reduction. PCA transforms the original variables into a set of uncorrelated principal components through linear transformation, where the first principal component explains the greatest variance in the data. In practical applications, the first principal component is chosen as the comprehensive centrality metric because it retains the majority of the original data information.

[0040] Before performing principal component analysis, the original centrality indices need to be standardized to eliminate dimensional differences between the different indices. After standardization, the covariance matrix or correlation coefficient matrix is ​​calculated and eigenvalue decomposition is performed on it to obtain eigenvalues ​​and eigenvectors. The eigenvector corresponding to the largest eigenvalue is selected as the projection direction, and the original data is projected onto this direction to obtain the composite centrality index.

[0041] After obtaining the comprehensive centrality index for each node, we average the comprehensive centrality index of the nodes within each subgraph to obtain the core node density of that subgraph. The core node density reflects the concentration of core nodes within a subgraph. A higher density indicates a denser core node population within the subgraph and a more compact subgraph structure.

[0042] To calculate the edge weight distribution, we traverse all edges in the subgraph, add up the weight of each edge, and then divide the sum by the number of edges to obtain the average edge weight. The average edge weight reflects the average strength of the connections between nodes in the subgraph. Higher weights indicate stronger connections between nodes in the subgraph.

[0043] Arrange the core node density and average edge weight of each subgraph in the order of hierarchical partitioning to generate a density-weight joint matrix. In the process of arrangement, strictly follow the order of subgraphs in hierarchical partitioning to ensure that the rows in the matrix are consistent with the subgraphs. Figure 1The first column of the matrix stores the core node density, and the second column stores the average edge weight.

[0044] After generating the joint density-weight matrix, it is converted into a one-dimensional topological structure vector through vectorization. Vectorization involves expanding a matrix into a one-dimensional array using either row-major or column-major order. In practical applications, the appropriate expansion method is selected based on subsequent processing needs. The vectorized topological structure vector retains all information in the original matrix while converting the multidimensional data into a one-dimensional form, facilitating subsequent calculations and processing.

[0045] To enhance the expressive power of the topology vector, we can also include subgraph hierarchical encoding information in the vector. Hierarchical encoding represents the position of the subgraph within the hierarchical partitioning. For example, we can use path encoding to convert the hierarchical path of the subgraph into a digital encoding. The hierarchical encoding is used as a dimension of the vector, along with the core node density and average edge weight to form a complete topology vector.

[0046] The topology vector generated in this way comprehensively considers the core node density and edge weight distribution of the subgraph, while preserving the hierarchical relationships between subgraphs. This vector representation can comprehensively and accurately characterize the topological structure characteristics of information resources, providing strong support for subsequent information matching and management. In practical applications, the topology vector can be combined with other feature vectors (such as timeliness feature vectors and semantic relevance feature vectors) to participate in information matching operations, improving matching accuracy and efficiency.

[0047] Furthermore, this method is highly scalable. If more topological features need to be considered, corresponding metrics, such as the clustering coefficient and eigenvector centrality, can be added to the calculation process. These metrics can then be combined into one or more comprehensive indices through principal component analysis and added to the topological structure vector. By continuously enriching and refining the representation of the topological structure vector, we can better reflect the inherent structural characteristics of information resources, thereby improving the performance and user experience of the information matching management system.

[0048] When dealing with large-scale networks, computational efficiency is also a consideration. To improve efficiency, parallel computing techniques can be employed to partition the network into multiple subnetworks and compute the topological features of each subnetwork in parallel. Network sampling can also be employed to reduce computational effort while preserving the network's structural characteristics. These optimization measures improve computational efficiency while ensuring the accuracy of topological feature extraction, making the method more suitable for practical large-scale information matching and management systems.

[0049] In practical applications, the topology vector generation method can be adjusted and optimized based on different business scenarios and needs. For example, for applications sensitive to local structure, the extraction and analysis of local subgraph features can be enhanced; for applications sensitive to global structure, the characterization of the overall network topology can be strengthened. By flexibly adjusting the method parameters and processes, the topology vector can be better adapted to different application needs, providing more precise support for information matching and management.

[0050] Example 3: In the dynamic matching module, a cross-modal alignment based on a heterogeneous attention mechanism is performed on the temporal feature vector and the topological structure vector to generate a temporal-topological association encoding.

[0051] Multi-resolution sampling is performed on the timeliness feature vector to generate hierarchical feature slices. The timeliness feature vector is generated by sliding-period decomposition and fusion of the real-time update frequency. It contains timeliness information, such as the short-term fluctuations and long-term trends of the information resource. Multi-resolution sampling uses a pyramid pooling method to downsample the feature vector at different resolutions along the time dimension. For example, sampling at resolutions of 1 / 2, 1 / 4, and 1 / 8 produces feature slices of varying granularity. During the sampling process, the sampling window size and step size at each resolution are adjusted accordingly. For higher-resolution sampling, the window size and step size are smaller, capturing more detailed timeliness changes. For lower-resolution sampling, the window size and step size are larger, extracting more macroscopic trend information. This multi-resolution sampling allows for comprehensive capture of the changing patterns of timeliness features at different time scales.

[0052] A tensor outer product operation is performed on each layer of slices and the topology vector to generate a multi-granularity interaction feature matrix. The topology vector is generated by extracting information such as core node density and edge weight distribution after subgraph partitioning. It reflects the topological characteristics of the information resource. The tensor outer product operation can combine temporal features and topology features to explore potential correlations between them. For each feature slice at each resolution, the outer product operation is performed on it with the topology vector to obtain a two-dimensional matrix. Each element of this matrix represents the interaction strength between temporal features and topology features at the corresponding location. The interaction feature matrices generated at different resolutions reflect the correlation patterns between temporal features and topology features at different time scales. For example, the interaction feature matrix at high resolution may capture the correlation between short-term temporal changes and local topology, while the interaction feature matrix at low resolution may reflect the relationship between long-term trends and global topology.

[0053] These matrices are stacked at different resolution levels to form a multi-granularity mapping tensor with the dimensions (number of layers × time dimension × topological dimension). This tensor contains the interactive information between temporal features and topological features at multiple resolutions, providing a rich data foundation for subsequent feature mining.

[0054] The kernel function's bandwidth parameter is dynamically selected based on the energy distribution of the multi-granularity mapping tensor. The energy distribution is determined by calculating the singular value decomposition (SVD) results for each dimension of the tensor. SVD decomposes the tensor into three matrices: a left singular matrix, a singular value matrix, and a right singular matrix. The diagonal elements of the singular value matrix represent the energy distribution of the tensor across each dimension. The bandwidth parameter is proportional to the square root of the singular value to accommodate feature interactions at different scales. This approach allows the kernel function's scope to be adaptively adjusted based on the feature energy distribution, giving greater attention to higher-energy dimensions and thus more effectively capturing important correlations between features.

[0055] Adaptive kernel function is used to perform nonlinear convolution operation on tensor to capture nonlinear correlation between features. The form of adaptive kernel function is ,in and is the input vector, is the bandwidth parameter, dynamically determined based on the singular value decomposition (SVD) described above. This kernel function maps nonlinear relationships in low-dimensional space to high-dimensional space, enabling linear classifiers to handle nonlinear problems. During the convolution operation, the kernel function slides across the tensor, calculating the similarity of local regions and extracting nonlinear correlation patterns between features. For example, certain temporal features may have complex nonlinear relationships with topological features, and this nonlinear convolution operation can effectively capture these relationships.

[0056] At the same time, skip connections are superimposed to preserve the original feature information and generate an optimized feature space. Skip connections are a technique in deep learning that allows the network to directly pass the input of certain layers to subsequent layers without passing through all intermediate layers. This method can avoid the vanishing gradient problem that occurs in deep networks while preserving the important information in the original features. In the process of generating the optimized feature space, the features processed by nonlinear convolution are added to the original features to form a comprehensive feature space that contains both the original information and the nonlinearly transformed information. This ensures that in subsequent processing, the feature association information mined by the nonlinear transformation can be utilized without losing important details in the original features.

[0057] The optimized feature space is input into the graph attention layer, and neighborhood node information is aggregated through a multi-head attention mechanism. The graph attention layer is a neural network layer specifically designed for processing graph-structured data. It can adaptively assign attention weights based on the strength of the associations between nodes. In the multi-head attention mechanism, each attention head independently calculates the attention weights between nodes. The weight calculation is based on the dot product and normalization operations of the node features. Specifically, for each node, its features are dot-producted with the features of other nodes to obtain a similarity score, which is then normalized using the softmax function to obtain the attention weight. Different attention heads focus on different feature subspaces, thereby capturing richer information about the relationships between nodes. For example, one attention head may focus on feature associations related to timeliness, while another attention head may focus on feature associations related to topology.

[0058] After propagation through a multi-layer graph neural network, global average pooling compresses feature dimensions to generate a temporal-topological correlation code. During the graph neural network propagation process, the features of each node are updated based on the features and attention weights of its neighboring nodes, allowing the node to obtain structural information from a wider range. Global average pooling averages the features of all nodes, compressing the feature representation of the graph structure into a vector of fixed dimension. This vector is the temporal-topological correlation code, which combines cross-modal correlation information between temporal and topological features, and can comprehensively reflect the characteristics of information resources in terms of both temporality and topology.

[0059] In practical applications, temporal-topological association coding can be used for a variety of tasks, such as information recommendation and resource matching. For example, in an information recommendation system, a temporal feature vector can be generated based on the user's historical behavior and current needs, while a topological structure vector can be generated based on the network structure of the information resource. After generating the association code using this method, the similarity between the user and the information resource is calculated to recommend the most relevant information to the user. This method, based on multi-dimensional feature fusion, can more accurately capture the matching relationship between user needs and information resources, improving the accuracy and personalization of recommendations.

[0060] Furthermore, this method is highly scalable. If more feature dimensions are needed, other types of feature vectors, such as semantic association feature vectors and matching probability feature vectors, can be incorporated into the association code generation process. These can then be fused with the temporal feature vectors and topological structure vectors using similar cross-modal alignment methods to generate more comprehensive association codes. By continuously enriching and refining feature fusion methods, we can better meet the needs of diverse application scenarios and enhance the performance and user experience of information matching management systems.

[0061] To improve computational efficiency when processing large-scale networks, batch processing and parallel computing techniques can be employed. Large networks can be divided into multiple small batches, and feature extraction and association code generation tasks for each batch are processed in parallel, with the results finally combined. Network sampling can also be performed to reduce computational effort while preserving the network's structural characteristics. These optimization measures can improve the system's processing power while maintaining association code quality, making it more suitable for practical large-scale information matching management systems.

[0062] In practical applications, the generation method for timeliness-topology correlation codes can be adjusted and optimized based on different business scenarios and needs. For example, for certain application scenarios with high timeliness requirements, the weight of timeliness feature vectors in the correlation code generation process can be increased; for certain application scenarios sensitive to topology, the mining and analysis of topological features can be strengthened. By flexibly adjusting method parameters and processes, timeliness-topology correlation codes can be better adapted to different application needs, providing more precise support for information matching and management.

[0063] The temporality-topology correlation encoding generated by this cross-modal alignment method based on the heterogeneous attention mechanism can effectively capture the complex correlation between temporality features and topological structure features, providing a more comprehensive and accurate feature representation for the information matching management system, thereby improving the accuracy and efficiency of information matching and providing users with a better service experience.

[0064] Example 4: Targeting the topological structure of information resources, a hierarchical subgraph partitioning and feature extraction method is employed to generate a topological structure vector that comprehensively reflects the network's structural characteristics. This method uses a community discovery algorithm to decompose complex networks into multiple levels of subcommunities. It then extracts the core node density and edge weight distribution characteristics of each subcommunity, ultimately forming a topological structure vector with a hierarchical encoding.

[0065] Construct a topological network model of information resources. This network uses information entities as nodes and the relationships between entities as edges. The edge weights represent the strength of the relationships. For example, in an academic literature network, nodes can represent entities such as papers, authors, and institutions, while edges can represent citations between papers and collaborations between authors. Edge weights can be determined based on factors such as the number of citations and the frequency of collaboration. After constructing the topological network, a hierarchical community discovery algorithm is used to partition the network. This algorithm decomposes the network into multiple levels of subcommunities, where nodes within each subcommunity are tightly connected, while connections between subcommunities are relatively sparse.

[0066] During the partitioning process, the Louvain algorithm is first used to perform a preliminary partitioning of the entire network, resulting in first-level subcommunities. The Louvain algorithm is a community discovery algorithm based on modularity optimization. It maximizes the network's modularity by iteratively merging nodes, thereby dividing the network into distinct communities. The Louvain algorithm is then applied again to each first-level subcommunity, yielding second-level subcommunities, and so on, until a preset hierarchical depth or subcommunity size threshold is reached. This creates a hierarchical community structure, in which each subcommunity is contained within a higher-level parent community.

[0067] Next, we calculate the core node density of each subgraph. For each subgraph, we first calculate a composite centrality metric for each node. This metric is derived by taking a weighted average of a node's degree centrality, betweenness centrality, and closeness centrality. Degree centrality measures the number of connections a node has, betweenness centrality measures the importance of a node as a bridge within the network, and closeness centrality measures the proximity of a node to other nodes in the network. By taking a weighted average of these three metrics, we obtain a comprehensive metric that fully reflects the importance of a node.

[0068] After calculating the comprehensive centrality index for each node, we average the comprehensive centrality index of the nodes in the entire subgraph to obtain the core node density of the subgraph. The core node density reflects the concentration of core nodes in the subgraph. A higher density indicates a denser core node population and a more compact subgraph structure.

[0069] Next, the edge weight distribution of each subgraph is calculated. This edge weight distribution is represented by the average edge weight. This is calculated by traversing all edges within the subgraph, summing the weight of each edge, and then dividing the sum by the number of edges to obtain the average edge weight. The average edge weight reflects the average strength of the connections between nodes within the subgraph; higher weights indicate stronger connections between nodes within the subgraph.

[0070] Arrange the core node density and average edge weight of each subgraph in the order of hierarchical partitioning to generate a density-weight joint matrix. In the process of arrangement, strictly follow the order of subgraphs in hierarchical partitioning to ensure that the rows in the matrix are consistent with the subgraphs. Figure 1 The first column of the matrix stores the core node density, and the second column stores the average edge weight.

[0071] To preserve the hierarchical relationships between subgraphs, each subgraph is hierarchically encoded. Hierarchical encoding uses path encoding to convert the subgraph's path within the hierarchical structure into a digital code. For example, if a subgraph is located in the second community on the first layer, the third community on the second layer, and the fourth community on the third layer, its hierarchical code can be represented as 2.3.4. The hierarchical code is used as a dimension of the vector, along with the core node density and average edge weight to form a complete topological structure vector.

[0072] In the process of generating topological structure vectors, an adaptive feature selection method is used, namely ,in represents the final topological structure vector, represents the core node density vector, represents the edge weight distribution vector, and is a learnable weight matrix, is the sigmoid activation function. This formula adaptively adjusts the importance of core node density and edge weight distribution in the final feature vector based on the characteristics of different subgraphs. This approach more accurately captures the structural characteristics of different subgraphs and improves the expressive power of the topological structure vector.

[0073] The generated topological structure vector has excellent hierarchical characteristics and expressive power. It not only reflects the core node density and edge weight distribution of each subgraph, but also preserves the hierarchical relationships between subgraphs through hierarchical encoding. This vector representation can comprehensively and accurately characterize the topological structure characteristics of information resources, providing strong support for subsequent information matching and management.

[0074] In practical applications, topological structure vectors can be combined with other feature vectors (such as timeliness feature vectors and semantic relevance feature vectors) to participate in information matching operations, improving matching accuracy and efficiency. For example, in an information recommendation system, timeliness feature vectors and semantic relevance feature vectors can be generated based on the user's historical behavior and current needs. Simultaneously, a topological structure vector can be generated based on the network structure of information resources. After these vectors are combined, the similarity between the user and the information resource is calculated to recommend the most relevant information to the user.

[0075] Furthermore, this method is highly scalable. If more topological features need to be considered, corresponding metrics, such as the clustering coefficient and eigenvector centrality, can be added to the calculation process. These metrics can then be combined into one or more comprehensive indices through principal component analysis and added to the topological structure vector. By continuously enriching and refining the representation of the topological structure vector, we can better reflect the inherent structural characteristics of information resources, thereby improving the performance and user experience of the information matching management system.

[0076] When dealing with large-scale networks, computational efficiency is also a consideration. To improve efficiency, parallel computing techniques can be employed to partition the network into multiple subnetworks and compute the topological features of each subnetwork in parallel. Network sampling can also be employed to reduce computational effort while preserving the network's structural characteristics. These optimization measures improve computational efficiency while ensuring the accuracy of topological feature extraction, making the method more suitable for practical large-scale information matching and management systems.

[0077] In practical applications, the topology vector generation method can be adjusted and optimized based on different business scenarios and needs. For example, for applications sensitive to local structure, the extraction and analysis of local subgraph features can be enhanced; for applications sensitive to global structure, the characterization of the overall network topology can be strengthened. By flexibly adjusting the method parameters and processes, the topology vector can be better adapted to different application needs, providing more precise support for information matching and management.

[0078] The topological structure vector generated by this hierarchical subgraph partitioning and feature extraction method can effectively capture the topological structure characteristics of information resources, provide a more comprehensive and accurate feature representation for the information matching management system, thereby improving the accuracy and efficiency of information matching and providing users with a better service experience.

[0079] Example 5: The process of generating the fitness score involves inputting the temporal-topological association code, the semantic-topological association code, and the probabilistic-topological association code into different models to calculate the contribution, and then generating the final score through linear combination. The following details the implementation of each step with specific examples: 1. Calculation of Timeliness Contribution For example, let's assume a user's current demand is "latest research progress in artificial intelligence." The system uses the collection module to collect behavioral data on recent frequent searches for keywords like "large language model training" and "multimodal AI," extracting active timeliness demand characteristics, such as a high level of interest in information updated within the past week. The attribute analysis module analyzes an academic paper's real-time update frequency as daily (a high-frequency update). The corresponding timeliness feature vector includes short-term fluctuations (daily peak update period) and long-term trends (a steady increase in update frequency over the past month) based on a sliding period decomposition.

[0080] The temporality feature vector and the topology structure vector (the core node density of the subgraph of the academic network in which this paper is based is 0.8, and the mean edge weight distribution is 0.6) are fed into a heterogeneous attention mechanism for cross-modal alignment, generating a temporality-topology correlation code. This code contains information about the association between update frequency and network structure at different time scales, such as the strength of connections between frequently updated nodes and other core nodes in the topological network.

[0081] The associated code is input into a Gaussian process regression model to calculate the contribution to timeliness. Based on the prior distribution of "update frequency - user click volume" in historical data (e.g., assuming that frequently updated resources are more likely to attract recently active users), the Gaussian process regression model calculates the degree of match between the code vector and the prior covariance matrix to output the contribution of the paper to timeliness. For example, if the model outputs a value of 0.75, it indicates that the paper's update frequency is highly compatible with the user's current timeliness requirements.

[0082] 2. Semantic Contribution Calculation For example, an academic paper contains keywords such as "Transformer architecture optimization" and "improved contextual understanding," forming a semantic association with the user requirement of "large language model training." The attribute parsing module calculates the paper's semantic association as 0.9 (based on context window analysis, local semantic similarity is 0.85, global topic consistency is 0.92, and both are normalized and concatenated into a feature vector).

[0083] The semantic association feature vectors are cross-modally aligned with the topological structure vectors (the core nodes of the subgraph containing the paper are highly cited scholars in the field) to generate a semantic-topological association encoding. This encoding captures the association between the semantic topics of the paper and the core nodes in the topological structure, for example, the strength of the semantic coupling between "Transformer"-related content and the core scholar nodes in the topological network.

[0084] The association encoding is input into the capsule network to calculate semantic contribution. The capsule network uses the encoding vector as a low-level capsule and iteratively updates the coupling coefficients of higher-level capsules through a dynamic routing mechanism. For example, low-level capsules represent semantic entities such as "Transformer architecture" and "contextual understanding," while higher-level capsules represent the comprehensive theme of "large language model training." After three routing iterations, the modulus of the high-level capsule output vector is 0.88, indicating a semantic contribution of 0.88, indicating a high degree of semantic alignment between the paper's content and user needs.

[0085] 3. Probability Contribution Calculation The historical match success rate for this academic paper is 80% (8 successful matches out of the past 10 recommendations). Exponential smoothing is used to predict the change in match probability over the next week, generating a match probability feature vector (predicted values ​​fluctuate between 0.78 and 0.82). This vector is then cross-modally aligned with the topological structure vector (the distribution of subgraph edge weights reflects the closeness of scholarly collaboration) to generate a probability-topology correlation code. This code contains information about the correlation between historical success rate trends and topological structure stability, such as the fact that subgraphs containing high-success-rate resources have minimal edge weight fluctuations.

[0086] The association code is input into a random forest classifier to calculate the probability contribution. The classifier ranks the "topology features - matching results" feature importance in the historical data (for example, the core node density has an importance weight of 0.4, and the average edge weight is 0.3). It then performs Shapley value decomposition on the code vector and calculates the contribution weight of each dimension. For example, the contribution weight of the core node density is 0.35, and the average edge weight is 0.28. After smoothing with an exponentially weighted moving average and compressing to the range of 0-1 using a sigmoid function, the output probability contribution is 0.81, indicating that the matching probability prediction based on historical data and topology is relatively reliable.

[0087] 4. Fitness Score Generation and Optimal Matching Assuming that the linear combination weights of timeliness contribution, semantic contribution, and probability contribution are 0.3, 0.4, and 0.3 respectively (determined through offline training data optimization, and the sum of the weights is 1), the fitness score of the academic paper is: The system also performs the same calculation process on other candidate information resources. For example, another paper has a timeliness contribution of 0.6 (update frequency is once a week), a semantic contribution of 0.95 (higher keyword matching), and a probability contribution of 0.75 (historical matching success rate fluctuates greatly). The fitness score is: Comparing the scores of all candidate resources, the first paper (0.822) has the highest score and is therefore selected as the optimal match. The system pushes this paper to the user and records the matching results to update the historical matching success rate database, providing data support for subsequent matching.

[0088] 5. Implementation Details and Extensions Model parameter initialization: The prior covariance matrix of Gaussian process regression is statistically generated based on historical data of similar requirements; the initial coupling coefficient of the capsule network is randomly initialized and optimized through backpropagation; the tree structure of the random forest classifier splits nodes based on the Gini index, and the number of trees is set to 100 to avoid overfitting.

[0089] Data preprocessing: All feature vectors are Z-score normalized before input into the model to ensure that all dimensions are of the same scale. For example, the short-term fluctuation component of the timeliness feature vector ranges from [-2, 2], and the semantic relevance feature vector ranges from [0, 1]. After normalization, both are converted to distributions with a mean of 0 and a standard deviation of 1.

[0090] Computational efficiency optimization: For large-scale candidate resource collections, batch processing technology is used to parallelize the contribution and fitness scores of each resource. For example, GPU-accelerated matrix operations reduced the scoring time for 100,000 papers from 2 hours to 30 minutes.

[0091] Dynamic Weight Adjustment: In real-time matching scenarios, the combined weights of various contribution factors can be dynamically adjusted based on real-time user feedback (such as clicks and favorites). For example, if the click-through rate of a recommendation result is lower than expected, the system automatically increases the weight of semantic contribution and decreases the weight of probabilistic contribution to prioritize matching resources with more precise semantics.

[0092] Terms are used solely to distinguish one entity or operation from another and do not necessarily require or imply any actual relationship or order between such entities or operations. Furthermore, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0093] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An information matching management system based on big data, characterized in that: include: The acquisition module is used to obtain user behavior data sets and extract the information demand characteristics that are currently active; Attribute parsing module, used to parse the real-time update frequency, semantic relevance, historical matching success rate and data topology characteristics of target information resources; A multidimensional modeling module is used to perform heterogeneous feature modeling on the real-time update frequency, semantic relevance, historical matching success rate and data topology structure features, respectively, to generate a timeliness feature vector, a semantic relevance feature vector, a matching probability feature vector and a topology structure vector; The dynamic matching module is used to perform heterogeneous space projection and feature coupling operations on the timeliness feature vector, the semantic association feature vector, and the matching probability feature vector with the topological structure vector to generate a fitness score for each information resource and select the resource with the highest score as the optimal matching object.

2. The information matching management system based on big data according to claim 1 is characterized in that: Heterogeneous feature modeling is performed on the real-time update frequency, semantic relevance, historical matching success rate and the data topology structure features respectively to generate a timeliness feature vector, a semantic relevance feature vector, a matching probability feature vector and a topology structure vector, including: performing sliding period decomposition on the real-time update frequency to extract short-term fluctuation and long-term trend components and fusing them to generate the timeliness feature vector; performing context window analysis on the semantic relevance to extract local semantic similarity and global topic consistency and standardizing and splicing them to generate the semantic relevance feature vector; performing exponential smoothing prediction on the historical matching success rate to generate a probability change trajectory as the matching probability feature vector; and performing subgraph partitioning on the data topology to extract core node density and edge weight distribution as the topology structure vector.

3. The information matching management system based on big data according to claim 2 is characterized in that: The data topology structure is sub-divided into subgraphs to extract the core node density and edge weight distribution, including: using a community discovery algorithm to hierarchically segment the topology network, then counting the node centrality index and edge weight cumulative value of each subgraph, and generating a density-weight joint matrix as the topology structure vector.

4. The information matching management system based on big data according to claim 1 is characterized in that: The temporality feature vector, the semantic association feature vector, and the matching probability feature vector are subjected to heterogeneous space projection and feature coupling operations with the topological structure vector to generate a fitness score for each information resource, including: performing cross-modal alignment based on a heterogeneous attention mechanism on the temporality feature vector and the topological structure vector to generate a temporality-topological association code; performing the cross-modal alignment based on a heterogeneous attention mechanism on the semantic association feature vector and the topological structure vector to generate a semantic-topological association code; performing the cross-modal alignment based on a heterogeneous attention mechanism on the matching probability feature vector and the topological structure vector to generate a probability-topological association code; calculating the temporality contribution, semantic contribution, and probability contribution respectively according to the temporality-topological association code, the semantic-topological association code, and the probability-topological association code; and generating the fitness score based on a linear combination of the temporality contribution, the semantic contribution, and the probability contribution.

5. The information matching management system based on big data according to claim 4 is characterized in that: Cross-modal alignment based on a heterogeneous attention mechanism is performed on the temporal feature vector and the topological structure vector to generate a temporal-topological association code, including: constructing a multi-granularity mapping tensor between the temporal feature vector and the topological structure vector, using an adaptive kernel function to perform feature transformation, and then generating the temporal-topological association code through a gated graph neural network.

6. The information matching management system based on big data according to claim 5 is characterized in that: Constructing a multi-granularity mapping tensor between the temporal feature vector and the topological structure vector, including: performing multi-resolution sampling on the temporal feature vector to generate layered feature slices, performing a tensor outer product operation on each layer of slices with the topological structure vector, and stacking them into the multi-granularity mapping tensor.

7. The information matching management system based on big data according to claim 5 is characterized in that: Adopting an adaptive kernel function to perform feature transformation includes: dynamically selecting a bandwidth parameter of the kernel function according to the energy distribution of the multi-granularity mapping tensor, performing a nonlinear convolution operation, and then superimposing a jump connection to generate an optimized feature space.

8. The information matching management system based on big data according to claim 5 is characterized in that: The temporality-topology association code is generated by a gated graph neural network, including: aggregating neighborhood node information from the optimized feature space input graph attention layer, and compressing it through a pooling operation after multi-layer propagation to generate the temporality-topology association code.

9. The information matching management system based on big data according to claim 4 is characterized in that: The timeliness contribution, semantic contribution and probability contribution are calculated respectively according to the timeliness-topology association code, the semantic-topology association code and the probability-topology association code, including: inputting the timeliness-topology association code into a Gaussian process regression model, and calculating the timeliness contribution according to a prior covariance matrix; inputting the semantic-topology association code into a capsule network, and iteratively generating the semantic contribution through a dynamic routing mechanism; inputting the probability-topology association code into a random forest classifier, and outputting the probability contribution based on feature importance sorting.

10. The information matching management system based on big data according to claim 9 is characterized in that: The output of the probability contribution based on the feature importance sorting includes: performing Shapley value decomposition on the probability-topology association code to calculate the contribution weight of each dimension, smoothing through exponentially weighted moving average, and finally using S-type function compression to generate a probability contribution in the range of 0 to 1.

Images