An adaptive path matching method and system based on multi-source heterogeneous knowledge graph

Abstract

The application relates to the technical field of knowledge graph recognition, and particularly discloses a self-adaptive path matching method and system based on a multi-source heterogeneous knowledge graph. The application integrates multi-source heterogeneous data, extracts static risk characteristic values, dynamic interaction characteristics and collaborative potential behavior characteristic values in steps, realizes multi-dimensional information depth coupling through tensor fusion, completes self-adaptive path recommendation based on similarity matching, adapts structured and unstructured data in data processing, considers static stability and dynamic evolution in feature extraction, simultaneously mines collaborative correlation characteristics, combines a preset knowledge graph library and a similarity algorithm in path matching, improves recommendation accuracy and pertinence, and can analyze and match a path through which a risk intention is transmitted and evolved in association with entities in a network from a global knowledge graph perspective, and thus can effectively predict a potential collaborative risk path problem.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to an adaptive path matching method and system based on multi-source heterogeneous knowledge graphs. Background Technology

[0002] With the rapid development of information technology and the increasing popularity of internet applications, massive amounts of multi-source, heterogeneous text data have emerged on various public and private online platforms. This data encompasses various forms such as social media updates, forum discussions, news, chat logs, emails, and transaction reports, potentially containing clues about the behavioral intentions, social relationships, activity patterns, and even potential illegal risks of individuals or organizations. Especially when issuing warnings and identifying potentially dangerous activities, effectively filtering, integrating, and analyzing the behavioral characteristics of target individuals from massive amounts of complex online information has become a key technological challenge in the fields of public safety and cyberspace governance. However, existing risk identification and intent analysis methods mostly rely on manually preset rules or limited labeled data (such as keyword scanning of chat logs from specific communication tools or anomaly detection of financial transaction data). They are unable to cope with the semantic diversity and pattern complexity brought about by the dynamic evolution of risk tactics. Secondly, analyzing a single type of data in isolation (such as analyzing only text or only transactions) will result in risk scores for isolated events or nodes. The lack of a global knowledge graph perspective makes it impossible to analyze and match the path of how risk intent spreads and evolves through entity associations in the network, thus leading to the inability to effectively predict potential collaborative risk paths. Therefore, an adaptive path matching method based on multi-source heterogeneous knowledge graphs is needed to solve the above problems. Summary of the Invention

[0003] The purpose of this invention is to provide an adaptive path matching method based on multi-source heterogeneous knowledge graphs, comprising: An adaptive path matching method based on multi-source heterogeneous knowledge graphs includes: Obtain a multi-source heterogeneous knowledge graph of the target object, wherein the multi-source heterogeneous knowledge graph includes structured text data and unstructured text data; Based on the structured text data, obtain the text intent nodes and text semantic constraint edges of the target object, and obtain the static risk feature values ​​based on the text intent nodes and text semantic constraint edges; Based on the unstructured text data, obtain the interactive text of the target object within a preset time sequence, and obtain the dynamic interaction density of multiple nodes based on the interactive text; Based on the text intent node and the dynamic interaction density of the nodes, multiple collaborative potential behavior feature values ​​are obtained, and the multiple collaborative potential behavior feature values, the static risk feature values ​​and the dynamic interaction density of the nodes are tensor fused to obtain a risk collaboration tensor with multidimensional risk association. Adaptive path matching is performed on the risk collaboration tensor based on a preset knowledge graph database to obtain a set of matching paths, and path recommendation results are generated based on the set of matching paths.

[0004] Preferably, the step of obtaining the text intent nodes and text semantic constraint edges of the target object based on the structured text data, and obtaining the static risk feature value based on the text intent nodes and the text semantic constraint edges, includes: Entity recognition and relation extraction are performed on the structured text data to obtain an initial knowledge subgraph, and the initial knowledge subgraph is divided according to a preset time node to obtain multiple knowledge subgraph nodes; Obtain the node co-occurrence frequency and semantic centrality of each knowledge subgraph node, and obtain the text intent node based on the node co-occurrence frequency and semantic centrality, and obtain the corresponding importance coefficient based on the text intent node; Obtain the path constraint strength and semantic relevance between each text intent node, and obtain the text semantic constraint edge based on the path constraint strength and semantic relevance, and obtain the corresponding constraint strength coefficient based on the text semantic constraint edge; The static risk contribution value of each knowledge subgraph node is calculated based on the importance coefficient and the constraint strength coefficient. The static risk contribution values ​​of all knowledge subgraph nodes are normalized to obtain multiple static risk contribution normalized values, and static risk feature values ​​are obtained based on the multiple static risk contribution normalized values.

[0005] Preferably, the step of obtaining the interaction text of the target object within a preset time sequence based on the unstructured text data, and obtaining the dynamic interaction density of multiple nodes based on the interaction text, includes: The preset time sequence is divided according to preset time nodes, resulting in multiple time nodes. The unstructured text data is segmented according to multiple time nodes to obtain interactive text in multiple time segments; Obtain the interaction entity and text semantics corresponding to each interaction text, and obtain the interaction behavior type based on the interaction entity and text semantics; Obtain the interaction frequency and intensity of each interaction behavior type in different time segments, and construct a node-time interaction matrix based on multiple interaction frequencies and multiple interaction intensities; The interaction stability index is obtained based on the frequency of interaction. Based on the node-time interaction matrix, obtain the slope of the interaction density change for each interaction behavior type within a preset time sequence; The dynamic interaction density of multiple nodes is obtained based on the slope of multiple interaction density changes and the multiple interaction stability indices.

[0006] Preferably, the step of obtaining multiple collaborative potential behavior feature values ​​based on the text intent node and the dynamic interaction density of multiple nodes includes: The initial weight value of the corresponding node is obtained based on the text intent node, and the dynamic salience score of each node is obtained based on the initial weight value of the node and the dynamic interaction density of multiple nodes. The maximum dynamic significance score is obtained by filtering the multiple dynamic significance scores. Obtain the neighboring nodes in the preset knowledge graph that have a direct semantic relationship with the node corresponding to the node with the maximum dynamic saliency score, and obtain the dynamic interaction density of the first node corresponding to the neighboring node. Based on cosine similarity, the dynamic interaction density of the first node is matched with the dynamic interaction density of multiple nodes to obtain multiple similar node dynamic interaction densities, and the multiple similar node dynamic interaction densities are used as multiple collaborative potential behavior feature values.

[0007] Preferably, the step of tensor fusion of multiple collaborative potential behavior feature values, static risk feature values, and multiple node dynamic interaction densities to obtain a multidimensional risk association risk collaboration tensor includes: Obtain the static risk feature value vector based on the static risk feature values; Obtain multiple node dynamic interaction density vectors based on the dynamic interaction densities of multiple nodes. Multiple collaborative latent behavior feature values ​​are used to obtain multiple collaborative latent behavior feature value vectors; The static risk feature value vector, the multiple node dynamic interaction density vectors, and the multiple collaborative potential behavior feature value vectors are tensor-fused and concatenated to obtain an initial fused tensor. The initial fusion tensor is subjected to tensor dimensionality reduction and normalization to obtain a risk collaboration tensor with multidimensional risk association.

[0008] Preferably, the step of adaptively matching the risk collaboration tensor based on a preset knowledge graph database to obtain a set of matching paths, and generating path recommendation results based on the set of matching paths, includes: In a pre-defined knowledge graph database, a set of candidate paths matching the risk collaboration tensor is obtained based on a cosine similarity algorithm, wherein the set of candidate paths includes multiple candidate paths; Obtain the path feature vector of each candidate path, and then perform weighted calculations on the risk collaboration tensor with multiple path feature vectors in sequence to obtain multiple comprehensive path feature values, and use the multiple comprehensive path feature values ​​as multiple path matching scores; Each candidate path is sorted according to the path matching score, and the top K paths with the highest scores are selected as the matching path set. The matching path set is then output as the path recommendation result.

[0009] This application also provides an adaptive path matching system based on a multi-source heterogeneous knowledge graph, including: The first acquisition module is used to acquire the multi-source heterogeneous knowledge graph of the target object, wherein the multi-source heterogeneous knowledge graph includes structured text data and unstructured text data; The second acquisition module is used to acquire the text intent nodes and text semantic constraint edges of the target object based on the structured text data, and to acquire static risk feature values ​​based on the text intent nodes and the text semantic constraint edges. The third acquisition module is used to acquire the interactive text of the target object within a preset time sequence based on the unstructured text data, and to acquire the dynamic interaction density of multiple nodes based on the interactive text. The fourth acquisition module is used to acquire multiple collaborative potential behavior feature values ​​based on the text intent node and the multiple node dynamic interaction densities, and to perform tensor fusion of the multiple collaborative potential behavior feature values, the static risk feature values ​​and the multiple node dynamic interaction densities to obtain a risk collaboration tensor of multidimensional risk association; The path matching module is used to perform adaptive path matching on the risk collaboration tensor based on a preset knowledge graph library, obtain a set of matching paths, and generate path recommendation results based on the set of matching paths.

[0010] Preferably, the second acquisition module includes: The first acquisition unit is used to perform entity recognition and relation extraction on the structured text data to obtain an initial knowledge subgraph, and to divide the initial knowledge subgraph according to a preset time node to obtain multiple knowledge subgraph nodes; The second acquisition unit is used to acquire the node co-occurrence frequency and semantic centrality of each knowledge subgraph node, and to acquire the text intent node based on the node co-occurrence frequency and the semantic centrality, and to acquire the corresponding importance coefficient based on the text intent node; The third acquisition unit is used to acquire the path constraint strength and semantic relevance between each text intent node, and to acquire the text semantic constraint edge based on the path constraint strength and the semantic relevance, and to acquire the corresponding constraint strength coefficient based on the text semantic constraint edge. The calculation unit is used to calculate the static risk contribution value of each knowledge subgraph node based on the importance coefficient and the constraint strength coefficient. The normalization processing unit is used to normalize the static risk contribution values ​​of all knowledge subgraph nodes to obtain multiple static risk contribution normalization values, and obtain static risk feature values ​​based on the multiple static risk contribution normalization values.

[0011] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0012] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0013] The beneficial effects of this application are as follows: Compared with traditional path recognition methods that rely on single data types, manual rules, or local feature analysis, this invention integrates multi-source heterogeneous data, extracts static risk feature values, dynamic interaction features, and collaborative potential behavior feature values ​​step by step, and achieves deep coupling of multi-dimensional information through tensor fusion. Finally, it completes adaptive path recommendation based on similarity matching, which significantly improves the recognition accuracy and adaptability of complex dynamic risk behaviors. At the data processing level, it achieves effective adaptation of structured and unstructured data. At the feature extraction level, it takes into account both static stability and dynamic evolution, and at the same time, it mines collaborative association features. At the path matching level, by combining a pre-set knowledge graph library and a similarity algorithm, it improves the accuracy and targeting of path recommendation. It can be effectively applied to risk paths in scenarios such as public safety and financial risk prevention and control that require the identification of potential behavioral paths. At the same time, from the perspective of a global knowledge graph, it can analyze and match the path of how risk intentions are propagated and evolved through entity associations in the network, thereby effectively predicting potential collaborative risk paths. Attached Figure Description

[0014] Fig. 1 This is a schematic diagram of a method flow according to an embodiment of this application.

[0015] Fig. 2 This is a schematic diagram of the system structure according to an embodiment of this application.

[0016] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0018] like Figs. 1-2 As shown, this application provides an adaptive path matching method based on multi-source heterogeneous knowledge graphs, including: An adaptive path matching method based on multi-source heterogeneous knowledge graphs includes: S1. Obtain the multi-source heterogeneous knowledge graph of the target object, wherein the multi-source heterogeneous knowledge graph includes structured text data and unstructured text data; S2. Obtain the text intent node and text semantic constraint edge of the target object based on the structured text data, and obtain the static risk feature value based on the text intent node and the text semantic constraint edge; S3. Obtain the interactive text of the target object within a preset time sequence based on the unstructured text data, and obtain the dynamic interaction density of multiple nodes based on the interactive text. S4. Obtain multiple collaborative potential behavior feature values ​​based on the text intent node and the multiple node dynamic interaction densities, and perform tensor fusion of the multiple collaborative potential behavior feature values, the static risk feature values ​​and the multiple node dynamic interaction densities to obtain a risk collaboration tensor of multidimensional risk association. S5. Based on a preset knowledge graph database, perform adaptive path matching on the risk collaboration tensor to obtain a set of matching paths, and generate path recommendation results based on the set of matching paths.

[0019] As described in steps S1-S5 above, with the development of network technology, massive amounts of multi-source heterogeneous text data implicitly contain the behavioral intentions, relationships, and potential risk clues of target objects. However, these data are scattered, semantically complex, and dynamically changing, making it difficult for traditional methods to comprehensively capture the related information. Existing technologies often rely on manually preset rules or limited labeled data, and can only perform local analysis on single types of data. They cannot uncover the propagation and evolution path of risk intentions through entity relationships from a global perspective, resulting in low accuracy and poor adaptability in recognizing dynamically changing collaborative behaviors. Therefore, this invention first obtains a multi-source heterogeneous knowledge graph of the target object (such as a specific account, person, or organization). The structured text data comes from various public or authorized database records, such as business registration information, judicial judgment documents, administrative penalty records, financial transaction reports, etc., which have clear fields and relational structures. The unstructured text data comes from the temporal interactive content generated by the target object in cyberspace, such as social media posts and comments, forum discussion posts, instant messaging group chat records, email text, etc., which are free in form, semantically rich, but loosely structured. By constructing a knowledge graph that integrates the two types of data, a unified and comprehensive data foundation is provided for subsequent in-depth analysis; The specific implementation process is as follows: Structured text data is obtained through pre-defined API interfaces or database queries. Entity recognition and relation extraction technologies (such as BERT-based NER models and predefined relation extraction rules) are used to automatically extract names, organization names, locations, times, events, and their relationships, constructing an initial structured knowledge subgraph. Unstructured text data is collected through web crawlers (complying with Robots Exclusion Protocol and laws and regulations) or log systems. After data cleaning (removing irrelevant symbols, advertisements, and duplicate content), word segmentation, and part-of-speech tagging, entity recognition and relation extraction models (such as open relation extraction based on dependency parsing) are also used to extract entities and semantic relationships, and attributes such as timestamps, sentiment polarity, and topic tags are added. The entities extracted from the two types of data are fused through entity disambiguation and alignment technologies (such as entity linking based on word vector similarity) to ultimately form a multi-source heterogeneous knowledge graph with the target object as the core, containing its related entities (such as contacts, trading partners, activity locations, and discussion topics) and various relationships (such as transaction relationships, communication relationships, membership relationships, and co-occurrence relationships). Secondly, static risk feature values ​​are extracted from structured text data. This is because structured data has the characteristics of standardized format and clear entity relationships, which can reflect the relatively stable attributes and related features of the target object. These features do not change frequently within a certain period of time and are the core carriers of static risk. This step is achieved through four key steps: First, entity recognition and relationship extraction are performed on the structured text data. Named entity recognition algorithms (such as the BiLSTM-CRF model) are used to identify key entities in the data (such as people, accounts, locations, and items), and relationship extraction algorithms (such as the TransE model) are used to extract the relationships between entities (such as transfer relationships and cooperation relationships), forming an initial knowledge subgraph. Then, multiple knowledge subgraph nodes are divided according to preset time nodes (such as by day or by month) to achieve time-series decomposition of static data. Next, the node co-occurrence frequency (the number of times the same node appears in different structured data) and semantic centrality (the importance of the node in the knowledge subgraph, calculated by the PageRank algorithm) of each knowledge subgraph node are calculated. Combining these two indicators, text intent nodes that can reflect the core intent of the target object are selected. The importance coefficients of nodes are assigned based on their role in expressing intent (e.g., nodes directly associated with risky behaviors have an importance coefficient of 0.8, while those indirectly associated have an importance coefficient of 0.3). Then, the path constraint strength (the tightness of the association between nodes, calculated by path length and association frequency) and semantic relevance (the semantic similarity of nodes, calculated by the cosine similarity algorithm) between text intent nodes are analyzed to determine the text semantic constraint edges and assign constraint strength coefficients (e.g., the constraint strength coefficient for directly associated nodes is 0.9, while that for indirectly associated nodes is 0.4). Finally, the static risk contribution value of each knowledge subgraph node is obtained by weighting the importance coefficient and constraint strength coefficient (e.g., static risk contribution value = importance coefficient × importance coefficient weight value + constraint strength coefficient × constraint strength coefficient weight value). After normalization (mapping the values ​​to the [0,1] interval), the values ​​are integrated into static risk feature values. The technical effect of this step is to accurately capture the static risk associations of the target object. For example, in financial risk analysis, by extracting frequently transferred account nodes (text intent nodes) through information such as account transfer relationships and transaction amounts in structured data, the constraint strength between them and other accounts is calculated, and the obtained static risk feature value can reflect the potential risk basis of the account. Extracting dynamic interaction density from unstructured text data is crucial because unstructured text data directly reflects the real-time interactive behavior of target objects. These behaviors change dynamically over time, revealing the dynamic evolution of risky behaviors. Therefore, time series analysis is needed to capture their patterns of change. The implementation process is as follows: First, a preset time series (e.g., the last 3 months) is divided into multiple time segments according to preset time nodes (e.g., by week). The unstructured text data is then segmented by time segments to obtain interactive text for different time periods. Next, entity recognition algorithms are used to extract interactive entities (e.g., chat partners, discussion objects) from each interactive text. Semantic analysis algorithms (e.g., TF-IDF, BERT models) are used to parse the text semantics, and the interaction behavior type (e.g., information consultation, resource exchange, conspiracy negotiation, etc.) is determined by combining both. Then, the interaction frequency (number of times the behavior type occurs) and interaction intensity (calculated through text semantic sentiment and keyword weights, such as those related to...) are statistically analyzed within different time segments for each interaction behavior type. (For sensitive words, the interaction intensity is higher), construct a node-time interaction matrix (rows represent interactive entity nodes, columns represent time segments, and elements are the combination values ​​of interaction frequency and intensity for the corresponding time period); then calculate the interaction stability index (such as the reciprocal of the standard deviation, the smaller the fluctuation, the higher the stability index) based on the fluctuation of the interaction frequency; based on the node-time interaction matrix, fit the slope of the interaction density change of each interaction behavior type within a preset time sequence using a linear regression algorithm to reflect the increase or decrease trend of interaction behavior; finally, weightedly fuse the slope of the interaction density change and the interaction stability index (such as node dynamic interaction density = slope of change × slope of change weight + stability index × stability index weight) to obtain the dynamic interaction density of multiple nodes. The technical effect of this step is to dynamically capture the interaction behavior changes of the target object. For example, when analyzing network risks, through group chat records in unstructured text data, it can be found that the interaction frequency of keywords such as betting and settlement gradually increases within a certain period of time (the slope of change is positive), and the interaction stability is high (small fluctuation), corresponding to a higher node dynamic interaction density, which can promptly identify the active trend of gang activities; First, the initial weight value of each node is determined based on the importance coefficient of the text intent node. This weight value is then multiplied by the node's dynamic interaction density to obtain the dynamic saliency score of each node (nodes with higher importance and more active interactions receive higher scores). The most representative node is selected by filtering based on the maximum value. Then, adjacent nodes with direct semantic association in the preset knowledge graph (such as nodes with semantic similarity greater than 0.7) are found, and their corresponding first node dynamic interaction density is obtained. Next, based on the cosine similarity algorithm, the first node dynamic interaction density is matched with the dynamic interaction densities of all nodes, and similar node dynamic interaction densities with similarity higher than a preset threshold (such as 0.8) are selected as collaborative potential behavior feature values. This process can uncover associated nodes with similar dynamic behaviors to the core node and capture potential collaborative behaviors. In the tensor fusion section, the static risk feature values, the dynamic interaction density of each node, and the collaborative potential behavior feature values ​​are first converted into feature vectors of the same dimension (e.g., by normalizing the vectors to map the values ​​to the same dimensional space). Then, tensor product operations are used to fuse and concatenate the three types of feature vectors to form an initial fusion tensor (the dimension of the static feature vectors × the number of dynamic interaction density vectors × the number of collaborative feature vectors). Finally, the initial fusion tensor is dimensionality reduced and redundant information is removed using a tensor decomposition algorithm (e.g., PARAFAC decomposition), while also being normalized to obtain a risk collaboration tensor with multidimensional risk associations. The technical effect of this step is to break the isolation between static and dynamic features, and between single node features and collaborative features. Tensor fusion achieves deep coupling of multidimensional information. For example, when analyzing smuggling gangs, static risk feature values ​​reflect the fixed fund account associations of gang members, node dynamic interaction density reflects the recent level of communication activity, and collaborative potential behavior feature values ​​reflect associated personnel with similar communication patterns to core members. The risk collaboration tensor after the fusion of these three can comprehensively characterize the gang's static organizational structure and dynamic activity characteristics. Based on a pre-defined knowledge graph database (containing a large amount of graph data of known risky or normal behavioral paths, constructed through historical case analysis, expert rule definition, and training with public data), adaptive path matching is performed on the risk collaboration tensor. First, a cosine similarity algorithm is used to calculate the similarity between the risk collaboration tensor and the path feature tensors in the pre-defined knowledge graph database, filtering out paths with similarity higher than a pre-defined threshold (e.g., 0.75) as a candidate path set. Then, the path feature vectors of each candidate path (e.g., path length, node association strength, behavioral type combination, etc.) are extracted, and the risk collaboration tensor and each path feature vector are weighted (the weights are set according to the importance and credibility of the path; for example, the weight of a path corresponding to a high-risk case is 0.9) to obtain the comprehensive path feature value of each candidate path, i.e., the path matching score. Finally, the paths are sorted from highest to lowest according to their matching scores, and the top K paths with the highest scores (K can be set according to actual needs, e.g., K=5) are selected as the matching path set, and the output is the path recommendation result. This enables precise matching and recommendation of potential behavioral paths for target objects, and can adapt to different types of risk behavior patterns. For example, if known dangerous paths (fund transfer, transportation, transaction connection) are stored in a pre-set knowledge graph library, when analyzing the risk collaboration tensor of a target object, if its matching score with the path is the highest, then that path is used as the recommendation result, providing a clear behavioral trajectory reference for risk warning. This approach can start from a global knowledge graph perspective and analyze and match the paths through which risk intentions propagate and evolve via entity associations in the network, thereby effectively predicting potential collaborative risk paths.

[0020] In one embodiment, step S2, which involves obtaining the text intent nodes and text semantic constraint edges of the target object based on the structured text data, and obtaining static risk feature values ​​based on the text intent nodes and the text semantic constraint edges, includes: S21. Perform entity recognition and relation extraction on the structured text data to obtain an initial knowledge subgraph, and divide the initial knowledge subgraph according to a preset time node to obtain multiple knowledge subgraph nodes; S22. Obtain the node co-occurrence frequency and semantic centrality of each knowledge subgraph node, and obtain the text intent node based on the node co-occurrence frequency and semantic centrality, and obtain the corresponding importance coefficient based on the text intent node; S23. Obtain the path constraint strength and semantic relevance between each text intent node, and obtain the text semantic constraint edge based on the path constraint strength and semantic relevance, and obtain the corresponding constraint strength coefficient based on the text semantic constraint edge; S24. Calculate the static risk contribution value of each knowledge subgraph node based on the importance coefficient and the constraint strength coefficient; S25. Normalize the static risk contribution values ​​of all knowledge subgraph nodes to obtain multiple static risk contribution normalization values, and obtain static risk feature values ​​based on the multiple static risk contribution normalization values.

[0021] As described in steps S21-S25 above, this invention transforms the original structured text data into a structured knowledge subgraph and performs time-series decomposition. The structured text data originates from data sources with fixed formats and defined fields, such as financial transaction reports, enterprise registration forms, business processing records, and logistics receipt ledgers, and can be obtained through database export, data interface calls, and offline document digitization. In terms of technical implementation, the structured text data is first processed using entity recognition algorithms. The BiLSTM-CRF model is selected as the core recognition framework. This model captures the semantic information of the text context through a bidirectional long short-term memory network (BiLSTM) and optimizes the entity boundary labeling accuracy by combining it with a conditional random field (CRF) model. It can accurately identify key entities in the data, including core elements such as people, accounts, company names, geographical locations, and item numbers. Subsequently, the TransE model is used for relation extraction. This model maps entities and relations to a low-dimensional vector space. By calculating the translation distance between entity vectors and relation vectors, the association between entities is determined, such as transfer relationships, equity relationships, agency relationships, etc. Based on the identified entities and extracted relations, an initial knowledge subgraph is constructed to intuitively present the association structure between entities and relations. Considering that while static data possesses stability, the association patterns may differ across different time periods, the initial knowledge graph needs to be divided according to preset time nodes (such as quarterly or annually; the specific time nodes can be set according to the risk evolution cycle of the application scenario, for example, financial risk scenarios can be divided by quarter, and corporate compliance scenarios can be divided by year). This results in multiple knowledge graph nodes, each corresponding to an entity association structure within a time segment, providing a foundation for subsequent time-segmented quantification of static risk feature values. The technical effect of this step is to transform unstructured text association information into a structured knowledge graph form. Simultaneously, the time-series division enables hierarchical processing of static data, avoiding feature ambiguity caused by cross-time-series data mixing. For example, when analyzing the cross-year related-party transaction risk of a company, the knowledge graph nodes divided by year can respectively present the transaction association structure of each year, facilitating the accurate capture of differences in static risk associations across different years. First, the co-occurrence frequency of each knowledge subgraph node is calculated, counting the number of times the node appears in the same knowledge subgraph and associated structured text data. The more times it appears, the closer the node is to the behavioral intent of the target object. Simultaneously, the PageRank algorithm is used to calculate the semantic centrality of the nodes. This algorithm determines the coreness of a node in the knowledge subgraph by iteratively calculating the inbound weights of the nodes. The higher the importance and the greater the number of inbound nodes, the higher the semantic centrality of the node. A screening threshold is constructed by combining the node co-occurrence frequency and semantic centrality (e.g., setting the co-occurrence frequency threshold to 3 times and the semantic centrality threshold to 0.6; the specific thresholds can be optimized through training with sample data). Nodes whose two indicators both exceed the thresholds are identified as text intent nodes. Then, an importance coefficient is assigned to the text intent nodes based on their coreness. A judgment matrix is ​​constructed using the analytic hierarchy process (AHP). Domain experts are invited to conduct pairwise comparisons of the importance of text intent nodes in risk identification. After consistency testing, a weight vector is calculated to obtain the importance coefficient of each text intent node. The coefficient ranges from [0,1]. The importance coefficient of core intent nodes (such as fund transfer accounts in financial risk scenarios) approaches 1, while the coefficient of secondary intent nodes is relatively lower. The technical effect of this step is to accurately screen out key nodes that can reflect the core intent of the target object, and to provide a basis for subsequent risk contribution value calculation by quantifying the importance coefficient. This avoids interference from irrelevant or secondary nodes on risk characteristics. For example, in potential risk scenarios, nodes such as "transaction meeting place" and "fund settlement account" that frequently appear and are at the semantic core will be screened as text intent nodes and given a high importance coefficient, thereby highlighting the core static risk points. First, the path constraint strength between each text intent node is calculated. Based on path analysis of the knowledge subgraph, the number of direct association paths between nodes and the path length (the number of intermediate nodes required for association between nodes) are counted. The more paths and the shorter the length, the higher the path constraint strength. The formula "Path constraint strength = Number of paths / (Average path length + 1)" is used for quantitative calculation to avoid the problem of zero denominator when the path length is 0. At the same time, the semantic relevance between nodes is calculated using the cosine similarity algorithm. The semantic vector of each text intent node (trained by the Word2Vec model, which maps nodes to low-dimensional semantic vectors based on the contextual information of structured text data) is used for similarity calculation. The smaller the angle between the semantic vectors, the higher the semantic relevance. Combining the path constraint strength and semantic relevance, when both indicators exceed the preset thresholds (e.g., the path constraint strength threshold is 0.5 and the semantic relevance threshold is 0.6), the association is determined to be a text semantic constraint edge. Subsequently, a hierarchical analysis method similar to the importance coefficient was employed, inviting domain experts to evaluate the constraint strength of the semantic constraint edges in the text. Combining the quantitative results of path constraint strength and semantic relevance, a constraint strength coefficient was calculated, with values ​​ranging from [0,1]. Edges with higher constraint strength (such as direct fund transfer edges) have coefficients closer to 1. The technical effect of this step is to accurately identify key connection edges between text intent nodes and quantify the degree of connection through the constraint strength coefficient. This provides a quantitative basis for the subsequent calculation of risk contribution values. For example, in smuggling risk analysis, the direct transaction connection edge between "supply enterprises" and "sales channel enterprises" has high path constraint strength and semantic relevance, resulting in a high constraint strength coefficient, which accurately reflects the contribution of this relationship to static risk. By fusing importance coefficients and constraint strength coefficients, the static risk contribution value of each knowledge subgraph node is obtained, achieving a preliminary quantification of static risk. Technically, a weighted summation method is used: Static Risk Contribution Value = Importance Coefficient × Importance Coefficient Weight Value + Constraint Strength Coefficient × Constraint Strength Coefficient Weight Value. The weight parameters need to be verified with sample data to ensure consistency between the quantification results and the actual risk situation. The technical effect of this step is that it integrates the quantitative indicators of node importance and edge constraint strength into a single risk contribution value, achieving a preliminary quantification of static risk and making the risk contribution level of each knowledge subgraph node comparable. The static risk contribution values ​​of all knowledge subgraph nodes are normalized to obtain standardized static risk feature values, laying the foundation for subsequent multi-dimensional feature fusion. Since the static risk contribution values ​​of different knowledge subgraph nodes may be at different orders of magnitude, direct fusion will lead to an imbalance in feature weights. Therefore, the Min-Max normalization algorithm is required to process them, mapping all static risk contribution values ​​to the [0,1] interval. The normalization formula is: Static risk contribution normalized value = (static risk contribution value of this node - minimum static risk contribution value of all nodes) / (maximum static risk contribution value of all nodes - minimum static risk contribution value of all nodes). Subsequently, based on the normalized values ​​of static risk contributions of all knowledge subgraph nodes, a weighted average is used to obtain static risk feature values. The weights are the temporal weights corresponding to each knowledge subgraph node (set according to the importance of the temporal segment in the preset temporal sequence; for example, if the weight of a recent temporal segment is higher than that of a distant one, the weight can be set using a linear weighting method, with a weight of 0.8 for recent nodes and 0.2 for distant nodes). The static risk feature value is calculated as Σ(normalized static risk contribution value × temporal weight). The technical effect of this step is to standardize the static risk feature values, eliminating the dimensional differences caused by different nodes and different temporal segments, and providing a unified comparison standard and fusion basis for the static risk feature values. For example, after normalization, the normalized values ​​of static risk contributions of all knowledge subgraph nodes are in the [0,1] range. The static risk feature values ​​obtained by temporal weighting can comprehensively reflect the static risk level of the target object in the entire preset temporal sequence, providing a standardized input indicator for subsequent fusion with dynamic features and collaborative features.

[0022] In one embodiment, step S3, which involves obtaining the interaction text of the target object within a preset time sequence based on the unstructured text data and obtaining the dynamic interaction density of multiple nodes based on the interaction text, includes: S31. Divide the preset time sequence according to preset time nodes to obtain multiple time nodes. S32. The unstructured text data is segmented according to multiple time nodes to obtain interactive text in multiple time segments; S33. Obtain the interaction entity and text semantics corresponding to each of the interaction texts, and obtain the interaction behavior type based on the interaction entity and the text semantics; S34. Obtain the interaction frequency and interaction intensity of each interaction behavior type in different time segments, and construct a node-time interaction matrix based on the multiple interaction frequencies and multiple interaction intensities; S35. Obtain the interaction stability index based on the interaction frequency; S36. Based on the node-time interaction matrix, obtain the slope of the interaction density change for each interaction behavior type within a preset time sequence; S37. Obtain the dynamic interaction density of multiple nodes based on the slope of multiple interaction density changes and the multiple interaction stability indices.

[0023] As described in steps S31-S37 above, this invention rationally divides the preset time sequence to provide a time framework for the subsequent temporal processing of unstructured text data. The setting of the preset time sequence needs to be combined with the risk evolution cycle of the application scenario. For example, in short-term risk warning scenarios (such as real-time network warnings), the preset time sequence can be set to 7 days; in long-term risk tracking scenarios (such as gang activity monitoring), the preset time sequence can be set to 3 months. The division of time nodes needs to ensure that the dynamic changes in interactive behavior can be captured. It is usually divided by day or week. If the preset time sequence is 3 months, dividing by week will yield 12 time nodes. The physical significance of this step is to discretize the continuous time dimension into multiple analyzable time segments, laying the foundation for subsequent splitting of interactive text in different time periods and analyzing the temporal changes in interactive behavior, avoiding distortion in dynamic feature extraction due to ambiguity in the time dimension. For example, when analyzing the chat logs of a certain online gang, dividing the three-month preset time sequence into 12 time nodes by week can clearly show the changes in the frequency of interaction of gang members in different weeks, providing a basis for time division for subsequent discovery of the active periods of the activity. Unstructured text data is segmented temporally. Unstructured text data originates from unformatted text such as social media posts, chat logs, forum discussions, emails, and news comments. It can be obtained through web crawlers (such as the Scrapy framework), data interface calls (such as social media open APIs), and local file reading (such as chat log text files). Technically, the acquired unstructured text data is first timestamped. The time node corresponding to each piece of text data is determined by time information in the text (such as sending time, publishing time) or the timestamp recorded during data collection. Then, based on the multiple time nodes defined in S31, the tagged unstructured text data is segmented into time segments, with each time segment corresponding to a time node, resulting in interactive text across multiple time segments. The technical effect of this step is to transform massive amounts of disordered unstructured text data into a collection of interactive texts arranged in an ordered manner according to time segments, thereby realizing the temporal regularization of unstructured data. For example, the WeChat chat records of a suspect over three months can be divided into 12 time segments of interactive text by week, with the chat records of each week grouped separately, which facilitates subsequent analysis of the differences in interactive behavior between different weeks. Interactive entities and textual semantics are extracted from the interactive text of each time segment, and the types of interactive behaviors are classified. Interactive entities, as the main participants in the interactive behavior, and textual semantics, as the core content of the interactive behavior, together determine the essential attributes of the interactive behavior. In terms of technical implementation, the BERT model is used to process the interactive text of each time segment. This model, through pre-trained semantic understanding capabilities, can accurately identify interactive entities in the text (such as person names, nicknames, account IDs, etc.), and obtain textual semantic vectors through semantic encoding, reflecting the core meaning of the interactive content. Subsequently, based on the association relationship of interactive entities (such as one-to-one interaction, many-to-many interaction) and the textual semantic vector, cluster analysis is performed (using the K-means clustering algorithm, with the number of clusters preset according to the application scenario, such as 5 categories, including information consultation, resource exchange, conspiracy negotiation, daily communication, and other categories), interactive texts with similar entity associations and semantic features are grouped into the same category, which is determined as an interactive behavior type. The technical effect of this step is to classify and organize interactive texts, summarizing complex and diverse interactive behaviors into a limited number of types, which facilitates subsequent targeted analysis of the dynamic characteristics of different types of interactive behaviors. For example, in potential risk scenarios, through semantic analysis, interactive texts containing keywords such as "source of goods," "delivery location," and "price" are classified as "conspiracy negotiation" interactive behaviors, providing a basis for subsequent key monitoring of the dynamic changes of this type of behavior. By quantifying the interaction frequency and intensity of each interaction type in different time segments, a node-time interaction matrix is ​​constructed to achieve a structured presentation of dynamic features. Interaction frequency directly reflects the activity level of a certain type of interaction behavior, while interaction intensity reflects the closeness of the association between interaction behaviors. In terms of technical implementation, the interaction frequency of each interaction behavior type in each time segment is counted, i.e., the number of interaction texts of that type; the TF-IDF algorithm is used to calculate the keyword weight of each interaction text, and combined with the magnitude of the text semantic vector, the interaction intensity of each interaction text is obtained (interaction intensity = sum of keyword weights × semantic vector magnitude), and the average interaction intensity of that type of interaction text in the corresponding time segment is taken as the interaction intensity of that time segment; then, a node-time interaction matrix is ​​constructed, where the rows of the matrix represent the nodes corresponding to the interaction behavior type, the columns represent the time nodes, and the matrix elements are the "interaction frequency - interaction intensity" combination values ​​of the corresponding interaction behavior type at the corresponding time node (stored in two-dimensional vector form, such as [frequency value, intensity value]). The technical effect of this step is to transform the dynamic characteristics of interactive behavior into structured matrix data, providing a standardized data source for subsequent dynamic indicator calculation. For example, the interaction frequency of the "conspiracy negotiation" type of interactive behavior in the third week is 15 times, and the average interaction intensity is 0.8. Then the element at the corresponding position in the node-time interaction matrix is ​​[15, 0.8], which intuitively presents the dynamic characteristics of this type of behavior in this period. The interaction stability index is calculated based on interaction frequency. This index reflects the degree of fluctuation of a certain type of interaction behavior in different time segments. The smaller the fluctuation, the higher the stability, and vice versa. In terms of technical implementation, the coefficient of variation is used as the quantitative indicator of the interaction stability index. The coefficient of variation = (standard deviation of interaction frequency) / (mean of interaction frequency), where the standard deviation of interaction frequency reflects the dispersion of frequency in different time segments, and the mean reflects the overall level of frequency. To facilitate subsequent calculations, the coefficient of variation is normalized (mapped to the [0,1] interval) to obtain the final interaction stability index. The closer the index is to 0, the higher the stability, and the closer it is to 1, the lower the stability. The technical effect of this step is that it quantifies the stability characteristics of interactive behavior, and can distinguish between regular interactive behavior and occasional interactive behavior. For example, if the frequency of a certain type of interactive behavior in 12 time segments is 10, 12, 11, 13, 10, 12, 11, 13, 12, 11, 13, 12, its standard deviation is small and the normalized coefficient of variation is 0.05, indicating that this type of interactive behavior has high stability and may be a normalized interaction of the target object. On the other hand, if the frequency of a certain type of interactive behavior fluctuates greatly, the standard deviation is large, and the coefficient of variation is close to 1, it may be an occasional temporary interaction and has low reference value for risk identification. Based on the node-time interaction matrix, a linear regression algorithm is used to fit the slope of the interaction density change for each interaction behavior type within a preset time sequence. This slope reflects the temporal evolution trend of the interaction density. Technically, the interaction density of each interaction behavior type at each time node is first calculated: Interaction density = (interaction frequency × interaction intensity) / total number of interactive texts within that time segment, ensuring comparability of interaction densities across different time segments. Then, using time nodes as independent variables (quantized as 1, 2, ..., n, where n is the number of time nodes) and interaction density as the dependent variable, a univariate linear regression model Y = kX + b (where k is the slope and b is the intercept term) is constructed. The model parameters are solved using the least squares method. By calculating the covariance and variance of the time variable and interaction density, the slope k (k = covariance / variance) is obtained. k > 0 indicates that the interaction density increases with time, k < 0 indicates that the interaction density decreases with time, and k = 0 indicates that the interaction density is basically stable. The technical effect of this step is that it accurately captures the dynamic evolution trend of interactive behavior, providing trend indicators for the calculation of the dynamic interaction density of subsequent nodes. For example, the slope of the interaction density change of "conspiracy negotiation" interactive behavior is k=0.15, indicating that the activity level of this type of behavior increases by 15% every week, showing a clear risk escalation trend, and providing key dynamic basis for risk warning. By fusing the slope of interaction density change with the interaction stability index, the dynamic interaction density of multiple nodes is obtained, achieving a comprehensive quantification of dynamic trends and stability characteristics. Technically, a weighted summation fusion method is adopted: Node dynamic interaction density = Interaction density change slope × ω1 + (1 - Interaction stability index) × ω2, where ω1 and ω2 are weight parameters set according to the risk focus of the application scenario. For example, in scenarios focusing on risk evolution trends, ω1 is set to 0.6 and ω2 to 0.4; in scenarios focusing on the regularity of interaction behavior, ω1 is set to 0.4 and ω2 to 0.6. The weight parameters need to be verified through sample data to ensure that the fusion result accurately reflects the dynamic risk characteristics. The technical effect of this step is that it integrates the trend and stability characteristics of interactive behavior, forming a quantitative indicator that can comprehensively reflect the dynamic interaction pattern. For example, the slope of change of a certain "conspiracy negotiation" interactive behavior is k=0.15, the interaction stability index is 0.05, and if ω1=0.6 and ω2=0.4, then the node dynamic interaction density is 0.15×0.6+(1-0.05)×0.4=0.47. This value reflects both the increasing trend of the behavior and its high stability, and comprehensively reflects the dynamic risk characteristics of this type of behavior.

[0024] In one embodiment, step S4, which involves obtaining multiple collaborative potential behavior feature values ​​based on the text intent node and the dynamic interaction density of the multiple nodes, includes: S41. Obtain the corresponding initial weight value of the node according to the text intent node, and obtain the dynamic salience score of each node according to the initial weight value of the node and the dynamic interaction density of multiple nodes; S42. Filter the multiple dynamic significance scores by maximum value to obtain the maximum dynamic significance score. S43. Obtain the adjacent nodes in the preset knowledge graph that have a direct semantic relationship with the node corresponding to the node with the maximum dynamic saliency score, and obtain the dynamic interaction density of the first node corresponding to the adjacent node. S44. Based on cosine similarity, the dynamic interaction density of the first node is matched with the dynamic interaction density of multiple nodes to obtain multiple similar node dynamic interaction densities, and the multiple similar node dynamic interaction densities are used as multiple collaborative potential behavior feature values.

[0025] As described in steps S41-S44 above, this invention obtains the dynamic saliency score of each node by fusing the initial weight value of the text intent node with the node's dynamic interaction density, thereby quantifying the dynamic importance of the node. The initial weight value of the text intent node is derived from the importance coefficient obtained in step S22. This coefficient has already quantified the static importance of the text intent node through node co-occurrence frequency and semantic center. Here, this coefficient is directly used as the initial weight value of the node, ensuring the consistency of feature transmission. In terms of technical implementation, a weighted product method is used to calculate the dynamic saliency score: Dynamic saliency score = Initial weight value of node × Dynamic interaction density of node. This calculation method retains the core guiding position of the text intent node while incorporating the activity level and evolution trend of the node's dynamic interaction, enabling the score to comprehensively reflect the node's "static importance + dynamic activity". The technical effect of this step is to break the isolation between static weights and dynamic features. By quantifying the dynamic saliency score, it is possible to quickly identify nodes that have both core intent attributes and active performance in the current dynamic interaction scenario. For example, in financial risk analysis, the initial weight value of a certain text intent node (core account) is 0.9, and its node dynamic interaction density is 0.8. Then the dynamic saliency score is 0.9 × 0.8 = 0.72. This score can intuitively reflect the importance of the node in the dynamic scenario and provide a clear basis for the subsequent selection of core nodes. By filtering multiple dynamic saliency scores to find the maximum value, the core node with the most dynamic importance in the current scenario is accurately located. Technically, the dynamic saliency scores of all nodes are traversed, and the score with the highest value is selected as the maximum dynamic saliency score. The node corresponding to this maximum dynamic saliency score is the most crucial node in the current dynamic interaction process. The physical significance of this step lies in focusing on the core influencing node. Because this node possesses both the highest static importance and the most active dynamic interaction characteristics, its behavioral patterns have the most significant impact on other related nodes, making it a key entry point for uncovering potential collaborative behaviors. The technical effect of this step is that it enables rapid location of core nodes, avoiding the blindness of multi-node analysis. For example, in a potential risk scenario, after score filtering, the node corresponding to the maximum dynamic saliency score may be the core contact node of the group. The interaction behavior of this node will directly affect the interaction patterns of other group members. Using this node as the core for collaborative feature mining can significantly improve the targeting and efficiency of identifying potential collaborative behaviors. This process involves identifying directly semantically related neighboring nodes of core nodes and obtaining their corresponding first-node dynamic interaction density to provide a reference benchmark for subsequent collaborative feature matching. The pre-defined knowledge graph is a foundational graph library containing massive entities and semantic relationships. It was constructed through prior data accumulation, integration of public knowledge graphs, and annotation by domain experts, covering direct semantic association rules between different entities (e.g., semantic similarity exceeding a pre-defined threshold of 0.7 is considered a direct semantic association). Technically, based on the pre-defined knowledge graph, neighboring nodes of the core node corresponding to the highest dynamic saliency score are retrieved. Nodes with direct semantic associations to the core node (e.g., direct communication objects of the core contact node, financial transaction nodes, etc.) are selected, and the node dynamic interaction density corresponding to these neighboring nodes is extracted and used as the first-node dynamic interaction density. The technical effect of this step is to establish dynamic associations between core nodes and related nodes, obtaining benchmark data that reflects the dynamic characteristics of directly related nodes surrounding the core node. For example, the first-node dynamic interaction density of neighboring nodes (direct transaction objects) of the core contact node provides a clear matching benchmark for finding other nodes with similar interaction patterns. The cosine similarity algorithm is used to match the dynamic interaction density of the first node with the dynamic interaction densities of all nodes, thereby identifying the dynamic interaction densities of nodes with collaborative behavior characteristics as potential collaborative behavior feature values. The cosine similarity algorithm measures the similarity between two vectors by calculating the cosine of the angle between them, with a value ranging from -1 to 1. A value closer to 1 indicates a higher similarity, and this algorithm effectively quantifies the similarity between two dynamic interaction density vectors. Technically, the dynamic interaction density of the first node and multiple nodes are first converted into vectors of the same dimension (e.g., dynamic interaction density sequence vectors based on time series). Then, the similarity between each pair of vectors is calculated using the cosine similarity formula. A similarity threshold is set (e.g., 0.8, which can be optimized through training with domain sample data). Node dynamic interaction densities with similarity values ​​higher than the threshold are selected as similar node dynamic interaction densities. The nodes corresponding to these density values ​​have similar dynamic interaction patterns with the core node and adjacent nodes, thus constituting potential collaborative behavior feature values. The technical effect of this step is to accurately identify the dynamic features of nodes with collaborative behavior patterns. For example, if the cosine similarity between the dynamic interaction density vector of the adjacent nodes of the core contact node and the dynamic interaction density vector of a certain node is 0.85, which is higher than the preset threshold, then the dynamic interaction density of that node is determined as a potential collaborative behavior feature value, indicating that the node has collaborative interaction behavior with the core associated node, providing key features that can reflect collaborative risks for the subsequent construction of risk collaboration tensors.

[0026] In one embodiment, step S4, which involves tensor fusion of multiple collaborative potential behavior feature values, static risk feature values, and multiple node dynamic interaction densities to obtain a multidimensional risk association risk collaboration tensor, includes: S45. Obtain the static risk feature value vector based on the static risk feature values; S46. Obtain multiple node dynamic interaction density vectors based on the multiple node dynamic interaction densities. S47. Obtain multiple collaborative latent behavior feature value vectors from the multiple collaborative latent behavior feature values; S48. Tensor fusion and concatenation of the static risk feature value vector, the multiple node dynamic interaction density vectors and the multiple collaborative potential behavior feature value vectors are performed to obtain an initial fusion tensor. S49. Perform tensor dimensionality reduction and normalization on the initial fusion tensor to obtain a risk collaboration tensor with multidimensional risk association.

[0027] As described in steps S45-S49 above, this invention converts static risk feature values ​​into static risk feature value vectors, providing a unified vector input format for subsequent tensor fusion. The static risk feature values ​​originate from the output of step S25 and are quantified values ​​reflecting the overall static risk level of the target object after normalization. Technically, a feature vector mapping method is used to expand a single static risk feature value into a fixed-dimensional vector. The vector dimension is set according to the feature complexity of the application scenario, for example, 128 dimensions. The first element is the static risk feature value, and the remaining elements are supplemented through zero-padding or feature expansion based on domain prior knowledge, ensuring the uniformity and rationality of the vector dimension. The technical effect of this step is to convert a single-dimensional static risk feature value into a high-dimensional vector form, giving it the basis for fusion with other multi-dimensional feature vectors, while retaining the core information of the static risk feature value. For example, when the static risk feature value is 0.75, the first element of the converted 128-dimensional static risk feature value vector is 0.75, and the remaining elements are 0, ensuring that the core information is not lost and meeting the dimensional requirements of subsequent tensor fusion. The dynamic interaction density of multiple nodes is transformed into a dynamic interaction density vector, achieving vector normalization of dynamic features. The dynamic interaction density of multiple nodes originates from the output of step S37, with each value quantifying the dynamic interaction pattern of the corresponding node. Technically, for each node's dynamic interaction density, the same dimensional setting as the static risk feature value vector (e.g., 128 dimensions) is adopted. The node's dynamic interaction density is used as the core element of the vector, and the remaining elements are supplemented through zero-padding or dynamic feature expansion (e.g., combining auxiliary information such as the node's interaction frequency and intensity), ensuring that the dimension of each node's dynamic interaction density vector is consistent with the static risk feature value vector. For example, if a node's dynamic interaction density is 0.6, the first element of the corresponding 128-dimensional vector is 0.6, and the remaining elements are assigned values ​​based on auxiliary information such as the node's interaction stability index. This preserves the core feature of the dynamic interaction density while enriching the information dimension of the vector. The technical effect of this step is to convert the dynamic interaction density values ​​of multiple discrete nodes into a set of vectors of a unified dimension, so that dynamic features can be fused with static features in the same dimensional space, providing a standardized dynamic feature input for subsequent tensor fusion, and enhancing the expressive power of dynamic features through vector expansion. This process transforms multiple collaborative latent behavior feature values ​​into multiple collaborative latent behavior feature value vectors, achieving a standardized expression of collaborative features. These collaborative latent behavior feature values, derived from the output of step S44, are key quantitative indicators reflecting multi-node collaborative behavior. Technically, the vector dimension (128 dimensions) is maintained as described earlier, employing the same construction logic as the node dynamic interaction density vector. Each collaborative latent behavior feature value is used as the core element of its corresponding vector, with the remaining elements supplemented through zero-padding or expansion using collaborative association information (such as combining the node association strength corresponding to the collaborative feature). This ensures that the collaborative latent behavior feature value vector maintains the same dimension as the static risk feature value vector and the node dynamic interaction density vector. For example, if a collaborative latent behavior feature value is 0.85, the first element of the corresponding vector is 0.85, and the remaining elements are assigned values ​​based on the semantic association degree of the node corresponding to that collaborative feature. This allows the vector to simultaneously carry the core numerical value and association information of the collaborative feature. The technical effect of this step is to achieve vector standardization of collaborative latent behavior feature values, enabling them to be integrated into a unified fusion framework, providing a guarantee for the organic integration of the three types of features. Simultaneously, vector expansion further mines the association information of collaborative features, enhancing the expressive power of the features. An initial fusion tensor is obtained by fusing and concatenating static risk feature vectors, multiple node dynamic interaction density vectors, and multiple collaborative potential behavior feature vectors. Technically, a combination of tensor product operations and vector concatenation is used for fusion. First, the dimensional structure of the initial fusion tensor is determined as "static feature vector dimension × number of node dynamic interaction density vectors × number of collaborative potential behavior feature vectors". For example, if the static feature vector is 128-dimensional, there are 5 node dynamic interaction density vectors, and 3 collaborative potential behavior feature vectors, then the initial fusion tensor has dimensions of 128 × 5 × 3. Then, the three types of vectors are filled into the tensor according to this dimensional structure. The static risk feature vector serves as the basic dimension throughout the tensor, while the multiple node dynamic interaction density vectors and multiple collaborative potential behavior feature vectors serve as the constituent elements of the other two dimensions. Tensor product operations are used to achieve deep coupling of the three types of vectors, forming the initial fusion tensor. The technical effect of this step is to break the isolation of the three types of features. Through tensor fusion, a deep association between static features, dynamic features and collaborative features is achieved. The initial fusion tensor can simultaneously carry the core information of the three types of features and the relationship between features. For example, an element at a certain position in the tensor contains both the value of the static risk feature and the dynamic interaction density and collaborative behavior features of the corresponding node, providing an integrated feature carrier for comprehensively characterizing multidimensional risk association. In the optimization process of the initial fusion tensor, a standardized risk collaboration tensor is obtained through tensor dimensionality reduction and normalization. Although the initial fusion tensor integrates multi-dimensional features, it often suffers from excessively high dimensionality and uneven data distribution, which increases the computational cost of subsequent path matching and affects matching accuracy. Technically, tensor dimensionality reduction employs the PARAFAC decomposition algorithm. This algorithm effectively reduces the tensor dimensionality while preserving core feature information by decomposing a high-dimensional tensor into multiple low-dimensional factor matrices. During the decomposition process, the number of factors is set to 30 (this parameter can be optimized based on sample data to ensure no loss of core information), reducing the initial high-dimensional tensor to a low-dimensional tensor. Subsequently, the Min-Max normalization algorithm is used to map all elements of the dimensionality-reduced tensor to the [0,1] interval. The normalization formula is: Normalized value of tensor element = (Original value of the element - Minimum value of the tensor element) / (Maximum value of the tensor element - Minimum value of the tensor element), thus standardizing the tensor data. The technical effect of this step is that it reduces the feature dimension through dimensionality reduction, thereby reducing the computational power consumption of subsequent path matching and improving the efficiency of the algorithm. At the same time, normalization eliminates the dimensional differences between different features, making the tensor data distribution more uniform and avoiding interference from excessively large or small feature values ​​on the matching results. For example, the tensor of 128×5×3 before dimensionality reduction becomes 30×5×3 after PARAFAC decomposition. After normalization, all elements are in the range of [0,1], which not only preserves the core information of multidimensional risk association, but also has good standardization characteristics, providing a high-quality feature carrier for subsequent accurate adaptive path matching.

[0028] In one embodiment, step S5, which involves adaptive path matching of the risk collaboration tensor based on a preset knowledge graph database to obtain a set of matching paths, and generating path recommendation results based on the set of matching paths, includes: S51. In a preset knowledge graph database, a set of candidate paths matching the risk collaboration tensor is obtained based on the cosine similarity algorithm, wherein the set of candidate paths includes multiple candidate paths; S52. Obtain the path feature vector of each candidate path, and perform weighted calculations on the risk collaboration tensor with multiple path feature vectors in sequence to obtain multiple comprehensive path feature values, and use the multiple comprehensive path feature values ​​as multiple path matching scores. S53. Sort each candidate path according to the path matching score, select the top K paths with the highest scores as the matching path set, and output the matching path set as the path recommendation result.

[0029] As described in steps S51-S53 above, this invention uses a cosine similarity algorithm to select a set of candidate paths that initially match the risk collaboration tensor from a pre-defined knowledge graph database. The pre-defined knowledge graph database is constructed by integrating historical case data, domain expert experience rules, and research results on publicly available behavioral paths. Each path in the database contains a corresponding path feature tensor, which has the same dimensions as the risk collaboration tensor and represents a quantitative integration of the path's features across all dimensions. Technically, the cosine similarity algorithm measures the similarity between the risk collaboration tensor and the feature tensor of each path in the database by calculating the cosine of the angle between them. The formula is cosθ = (A・B) / (||A|| × ||B||), where A is the risk collaboration tensor, B is the path feature tensor, A・B is the dot product of the two, and ||A|| and ||B|| are their magnitudes, respectively. A similarity threshold is set (e.g., 0.65, which is optimized through training with sample data to ensure coverage of effective paths while reducing redundant candidates). Paths with cosine similarity higher than the threshold are then selected to form a candidate path set. The technical advantage of this step is that it quickly narrows down the matching range, selecting candidate paths with a certain correlation to the risk characteristics of the target object from a massive number of path templates. This avoids the computational consumption caused by matching all paths and improves matching efficiency. For example, when analyzing the risk collaboration tensor of a suspected behavior, cosine similarity calculation is used to select 30 paths with a similarity higher than 0.65 from 1000 paths in the database as a candidate path set, laying the foundation for subsequent accurate matching. A precise quantitative evaluation of the candidate path set is performed, and a path matching score for each candidate path is obtained through weighted calculation. The path feature vector of each candidate path is the result of dimensionality reduction processing of its path feature tensor, preserving the core feature information of the path, and its dimension is consistent with the feature dimension of the risk collaboration tensor. In terms of technical implementation, firstly, a weight coefficient is assigned to the path feature vector of each candidate path based on factors such as the path's risk level, frequency of occurrence, and credibility. For example, the weight coefficient of a high-risk path is 0.9, and the weight coefficient of a low-credibility path is 0.5. The weight coefficient is determined through the analytic hierarchy process combined with expert evaluation to ensure that the weight allocation is consistent with the risk focus of the actual application scenario. Then, the risk collaboration tensor and the path feature vector of each candidate path are multiplied by the corresponding weight coefficient to obtain the comprehensive path feature value, which is the path matching score. The calculation formula is: Path matching score = (Risk collaboration tensor ・ Path feature vector) × Weight coefficient. The technical effect of this step is that it achieves differentiated evaluation of candidate paths through weighted calculation. It considers both the matching degree between the risk collaboration tensor and the path features, as well as the importance and credibility of the path itself. For example, if the dot product between a candidate path and the risk collaboration tensor is 0.8 and its weight coefficient is 0.9 (high-risk path), then the path matching score is 0.8 × 0.9 = 0.72, which can accurately reflect the degree of fit between the path and the risk features of the target object and its own importance. By sorting and selecting the best paths based on their matching scores, the final set of matching paths is determined, and the path recommendation results are output. Technically, firstly, all paths in the candidate path set are sorted from highest to lowest matching score. A higher score indicates a higher degree of alignment with the risk characteristics of the target object and greater reference value. Then, a K value is set according to actual application needs (e.g., K=5; the K value can be adjusted according to the granularity of risk warning; for example, K value can be set to 3 in high-risk scenarios to balance accuracy and simplicity; K value can be set to 8 in complex scenarios to ensure coverage of multiple potential paths). The top K paths with the highest scores are selected to form the matching path set. Finally, this set is output in a structured form as the path recommendation result, including the specific process of the path, risk level, key nodes, and other information. The technical effect of this step is to accurately select the most representative core paths, providing a clear reference for risk decision-making. For example, after sorting the candidate path set, the top 5 paths with the highest scores are selected as the matching path set. The path ranked first has a matching score of 0.85, corresponding to the path of "account registration - false advertising - fund collection - withdrawal and transfer". This result can directly provide a clear behavioral trajectory reference for the risk disposal of relevant departments. In summary, a cosine similarity algorithm was used to achieve accurate initial screening of multi-dimensional features. Weighted calculations incorporated differentiated importance assessment of paths, and ranking and optimization ensured efficient selection of core paths, constructing a closed-loop matching process of "initial screening - precise calculation - optimization." This process effectively solves the problem of balancing accuracy and efficiency in existing path matching technologies. In application scenarios, the generated path recommendations accurately reflect the potential behavioral paths of target objects, providing direct and reliable decision support for risk warning and handling. Compared to existing technologies that rely solely on single similarity calculations or simple ranking, the path matching results from this step are more targeted and valuable, significantly improving the accuracy of risk identification and decision-making efficiency, laying a crucial foundation for the practical application of the overall adaptive path matching method.

[0030] This application also provides an adaptive path matching system based on a multi-source heterogeneous knowledge graph, including: The first acquisition module is used to acquire the multi-source heterogeneous knowledge graph of the target object, wherein the multi-source heterogeneous knowledge graph includes structured text data and unstructured text data; The second acquisition module is used to acquire the text intent nodes and text semantic constraint edges of the target object based on the structured text data, and to acquire static risk feature values ​​based on the text intent nodes and the text semantic constraint edges. The third acquisition module is used to acquire the interactive text of the target object within a preset time sequence based on the unstructured text data, and to acquire the dynamic interaction density of multiple nodes based on the interactive text. The fourth acquisition module is used to acquire multiple collaborative potential behavior feature values ​​based on the text intent node and the multiple node dynamic interaction densities, and to perform tensor fusion of the multiple collaborative potential behavior feature values, the static risk feature values ​​and the multiple node dynamic interaction densities to obtain a risk collaboration tensor of multidimensional risk association; The path matching module is used to perform adaptive path matching on the risk collaboration tensor based on a preset knowledge graph library, obtain a set of matching paths, and generate path recommendation results based on the set of matching paths.

[0031] In one embodiment, the second acquisition module includes: The first acquisition unit is used to perform entity recognition and relation extraction on the structured text data to obtain an initial knowledge subgraph, and to divide the initial knowledge subgraph according to a preset time node to obtain multiple knowledge subgraph nodes; The second acquisition unit is used to acquire the node co-occurrence frequency and semantic centrality of each knowledge subgraph node, and to acquire the text intent node based on the node co-occurrence frequency and the semantic centrality, and to acquire the corresponding importance coefficient based on the text intent node; The third acquisition unit is used to acquire the path constraint strength and semantic relevance between each text intent node, and to acquire the text semantic constraint edge based on the path constraint strength and the semantic relevance, and to acquire the corresponding constraint strength coefficient based on the text semantic constraint edge. The calculation unit is used to calculate the static risk contribution value of each knowledge subgraph node based on the importance coefficient and the constraint strength coefficient. The normalization processing unit is used to normalize the static risk contribution values ​​of all knowledge subgraph nodes to obtain multiple static risk contribution normalization values, and obtain static risk feature values ​​based on the multiple static risk contribution normalization values.

[0032] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0033] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0034] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0035] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0036] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An adaptive path matching method based on multi-source heterogeneous knowledge graphs, characterized in that, include: Obtain a multi-source heterogeneous knowledge graph of the target object, wherein the multi-source heterogeneous knowledge graph includes structured text data and unstructured text data; Based on the structured text data, obtain the text intent nodes and text semantic constraint edges of the target object, and obtain the static risk feature values ​​based on the text intent nodes and text semantic constraint edges; Based on the unstructured text data, the interactive text of the target object within a preset time sequence is obtained, and the dynamic interaction density of multiple nodes is obtained based on the interactive text, including: The preset time sequence is divided according to preset time nodes, resulting in multiple time nodes. The unstructured text data is segmented according to multiple time nodes to obtain interactive text in multiple time segments; Obtain the interaction entity and text semantics corresponding to each interaction text, and obtain the interaction behavior type based on the interaction entity and text semantics; Obtain the interaction frequency and intensity of each interaction behavior type in different time segments, and construct a node-time interaction matrix based on multiple interaction frequencies and multiple interaction intensities; The interaction stability index is obtained based on the frequency of interaction. Based on the node-time interaction matrix, obtain the slope of the interaction density change for each interaction behavior type within a preset time sequence; The dynamic interaction density of multiple nodes is obtained based on the slope of multiple interaction density changes and the multiple interaction stability indices. Multiple collaborative potential behavior feature values ​​are obtained based on the text intent node and the dynamic interaction density of multiple nodes. The multiple collaborative potential behavior feature values, the static risk feature values ​​and the dynamic interaction density of multiple nodes are then fused by tensor to obtain a risk collaboration tensor with multidimensional risk association. Adaptive path matching is performed on the risk collaboration tensor based on a preset knowledge graph database to obtain a set of matching paths, and path recommendation results are generated based on the set of matching paths.

2. The adaptive path matching method based on multi-source heterogeneous knowledge graphs according to claim 1, characterized in that, The step of obtaining the text intent nodes and text semantic constraint edges of the target object based on the structured text data, and obtaining the static risk feature value based on the text intent nodes and the text semantic constraint edges, includes: Entity recognition and relation extraction are performed on the structured text data to obtain an initial knowledge subgraph, and the initial knowledge subgraph is divided according to a preset time node to obtain multiple knowledge subgraph nodes; Obtain the node co-occurrence frequency and semantic centrality of each knowledge subgraph node, and obtain the text intent node based on the node co-occurrence frequency and semantic centrality, and obtain the corresponding importance coefficient based on the text intent node; Obtain the path constraint strength and semantic relevance between each text intent node, and obtain the text semantic constraint edge based on the path constraint strength and semantic relevance, and obtain the corresponding constraint strength coefficient based on the text semantic constraint edge; The static risk contribution value of each knowledge subgraph node is calculated based on the importance coefficient and the constraint strength coefficient. The static risk contribution values ​​of all knowledge subgraph nodes are normalized to obtain multiple static risk contribution normalized values, and static risk feature values ​​are obtained based on the multiple static risk contribution normalized values.

3. The adaptive path matching method based on multi-source heterogeneous knowledge graphs according to claim 1, characterized in that, The step of obtaining multiple collaborative potential behavior feature values ​​based on the text intent node and the dynamic interaction density of multiple nodes includes: The initial weight value of the corresponding node is obtained based on the text intent node, and the dynamic salience score of each node is obtained based on the initial weight value of the node and the dynamic interaction density of multiple nodes. The maximum dynamic significance score is obtained by filtering the multiple dynamic significance scores. Obtain the neighboring nodes in the preset knowledge graph that have a direct semantic relationship with the node corresponding to the node with the maximum dynamic saliency score, and obtain the dynamic interaction density of the first node corresponding to the neighboring node. Based on cosine similarity, the dynamic interaction density of the first node is matched with the dynamic interaction density of multiple nodes to obtain multiple similar node dynamic interaction densities, and the multiple similar node dynamic interaction densities are used as multiple collaborative potential behavior feature values.

4. The adaptive path matching method based on multi-source heterogeneous knowledge graphs according to claim 1, characterized in that, The step of tensor fusion of multiple collaborative potential behavioral feature values, static risk feature values, and multiple node dynamic interaction densities to obtain a multidimensional risk association risk collaboration tensor includes: Obtain the static risk feature value vector based on the static risk feature values; Obtain multiple node dynamic interaction density vectors based on the dynamic interaction densities of multiple nodes. Multiple collaborative latent behavior feature values ​​are used to obtain multiple collaborative latent behavior feature value vectors; The static risk feature value vector, the multiple node dynamic interaction density vectors, and the multiple collaborative potential behavior feature value vectors are tensor-fused and concatenated to obtain an initial fused tensor. The initial fusion tensor is subjected to tensor dimensionality reduction and normalization to obtain a risk collaboration tensor with multidimensional risk association.

5. The adaptive path matching method based on multi-source heterogeneous knowledge graphs according to claim 1, characterized in that, The step of performing adaptive path matching on the risk collaboration tensor based on a preset knowledge graph database to obtain a set of matching paths, and generating path recommendation results based on the set of matching paths, includes: In a pre-defined knowledge graph database, a set of candidate paths matching the risk collaboration tensor is obtained based on a cosine similarity algorithm, wherein the set of candidate paths includes multiple candidate paths; Obtain the path feature vector of each candidate path, and then perform weighted calculations on the risk collaboration tensor with multiple path feature vectors in sequence to obtain multiple comprehensive path feature values, and use the multiple comprehensive path feature values ​​as multiple path matching scores; Each candidate path is sorted according to the path matching score, and the top K paths with the highest scores are selected as the matching path set. The matching path set is then output as the path recommendation result.

6. An adaptive path matching system based on multi-source heterogeneous knowledge graphs, characterized in that, include: The first acquisition module is used to acquire the multi-source heterogeneous knowledge graph of the target object, wherein the multi-source heterogeneous knowledge graph includes structured text data and unstructured text data; The second acquisition module is used to acquire the text intent nodes and text semantic constraint edges of the target object based on the structured text data, and to acquire static risk feature values ​​based on the text intent nodes and the text semantic constraint edges. The third acquisition module is used to acquire the interaction text of the target object within a preset time sequence based on the unstructured text data, and to acquire the dynamic interaction density of multiple nodes based on the interaction text, including: The preset time sequence is divided according to preset time nodes, resulting in multiple time nodes. The unstructured text data is segmented according to multiple time nodes to obtain interactive text in multiple time segments; Obtain the interaction entity and text semantics corresponding to each interaction text, and obtain the interaction behavior type based on the interaction entity and text semantics; Obtain the interaction frequency and intensity of each interaction behavior type in different time segments, and construct a node-time interaction matrix based on multiple interaction frequencies and multiple interaction intensities; The interaction stability index is obtained based on the frequency of interaction. Based on the node-time interaction matrix, obtain the slope of the interaction density change for each interaction behavior type within a preset time sequence; The dynamic interaction density of multiple nodes is obtained based on the slope of multiple interaction density changes and the multiple interaction stability indices. The fourth acquisition module is used to acquire multiple collaborative potential behavior feature values ​​based on the text intent node and the multiple node dynamic interaction densities, and to perform tensor fusion of the multiple collaborative potential behavior feature values, the static risk feature values ​​and the multiple node dynamic interaction densities to obtain a risk collaboration tensor of multidimensional risk association. The path matching module is used to perform adaptive path matching on the risk collaboration tensor based on a preset knowledge graph library, obtain a set of matching paths, and generate path recommendation results based on the set of matching paths.

7. The adaptive path matching system based on a multi-source heterogeneous knowledge graph according to claim 6, characterized in that, The second acquisition module includes: The first acquisition unit is used to perform entity recognition and relation extraction on the structured text data to obtain an initial knowledge subgraph, and to divide the initial knowledge subgraph according to a preset time node to obtain multiple knowledge subgraph nodes; The second acquisition unit is used to acquire the node co-occurrence frequency and semantic centrality of each knowledge subgraph node, and to acquire the text intent node based on the node co-occurrence frequency and the semantic centrality, and to acquire the corresponding importance coefficient based on the text intent node; The third acquisition unit is used to acquire the path constraint strength and semantic relevance between each text intent node, and to acquire the text semantic constraint edge based on the path constraint strength and the semantic relevance, and to acquire the corresponding constraint strength coefficient based on the text semantic constraint edge. The calculation unit is used to calculate the static risk contribution value of each knowledge subgraph node based on the importance coefficient and the constraint strength coefficient. The normalization processing unit is used to normalize the static risk contribution values ​​of all knowledge subgraph nodes to obtain multiple static risk contribution normalization values, and obtain static risk feature values ​​based on the multiple static risk contribution normalization values.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images