Methods, apparatus, electronic devices, and computer programs for generating relational graphs

Abstract

This application provides a method, apparatus, electronic device, and computer program product for generating relationship graphs, belonging to the field of data processing, to solve the problems of existing methods being unable to depict complex relationships between target objects and effectively identify fraud risks. The method includes the following steps: acquiring textual and numerical information of the target objects; performing feature vectorization processing on the textual and numerical information to obtain target feature vectors; performing cluster analysis based on the target feature vectors to obtain several clusters; calculating the similarity between target feature vectors within each cluster; determining the connection relationships of the target feature vectors based on the similarity to generate a relationship graph; and correcting the relationship graph based on external feedback information.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method, apparatus, electronic device, and computer program product for generating relational graphs. Background Technology

[0002] In the field of fraud prevention auditing, especially for identifying organized fraud in insurance policies, existing methods suffer from two main limitations. First, at the technical implementation level, current anti-fraud methods based on relationship networks largely rely on single feature sources for correlation analysis. Furthermore, when vectorizing text, they typically only use the final output layer representation of a pre-trained language model, resulting in incomplete feature information, loss of contextual semantics, and difficulty in accurately depicting complex relationships between objects. Second, at the business application level, auditing work heavily relies on manually written rules to screen suspicious policies from relational databases. This approach has limited coverage, low efficiency, and can only reveal superficial rule matching relationships between policies, failing to automatically uncover deep, hidden organized fraud networks and thus failing to meet the needs for efficient and accurate auditing. Summary of the Invention

[0003] This application provides a method, apparatus, electronic device, and computer program product for generating relationship graphs, which can solve the problems that existing methods have difficulty in depicting complex relationships between target objects and in effectively identifying fraud risks.

[0004] In a first aspect, embodiments of this application provide a method for generating a relational graph, the method comprising the following steps: Obtain textual and numerical information of the target object, and perform feature vectorization processing on the textual and numerical information to obtain the target feature vector; Cluster analysis is performed based on the target feature vectors to obtain several clusters; Calculate the similarity between target feature vectors within each cluster, determine the connection relationship of the target feature vectors based on the similarity, and generate a relationship graph; The relationship graph is revised based on external feedback.

[0005] Secondly, embodiments of this application provide a relational graph generation apparatus, which includes the following: The feature generation module is used to acquire textual and numerical information of the target object, and to perform feature vectorization processing on the textual and numerical information to obtain the target feature vector. The clustering processing module is used to perform clustering analysis based on the target feature vector to obtain several clusters; A graph generation module is used to calculate the similarity between target feature vectors within each cluster, determine the connectivity relationships of the target feature vectors based on the similarity, and generate a relational graph; and... The feedback correction module corrects the relationship graph based on external feedback information.

[0006] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the relational graph generation method as described in the first aspect.

[0007] Fourthly, embodiments of this application provide a computer program product, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, the program instructions being executed by a computer to implement the steps of the relational graph generation method as described in the first aspect.

[0008] This application's embodiments construct target feature vectors by fusing multi-dimensional information from textual and numerical sources, overcoming the limitations of single feature sources and enabling more comprehensive object representation. The method of first clustering and then dividing to construct the graph maintains the local structure and continuity of data distribution, avoiding structural damage caused by random sampling. This allows the generated relationship graph to more accurately reflect the deep, multi-dimensional relationships between objects. Graph-based association retrieval reduces the workload of manual review, while clustering supports global or specific regional analysis, enabling auditing work to quickly focus on different scenario needs, thus expanding the coverage and flexibility of risk screening. Furthermore, introducing external feedback to correct the relationship graph allows it to continuously absorb business experience and adapt to changes in fraud methods, thereby improving the model's practicality and long-term effectiveness. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating a method for generating a relational graph according to an embodiment of this application; Figure 2 This is a flowchart illustrating another method for generating a relational graph provided in an embodiment of this application; Figure 3This is a schematic diagram of the structure of a relational graph generation device provided in an embodiment of this application; Figure 4 This is a flowchart illustrating another method for generating a relational graph provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0012] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] Currently, the main methods used in building anti-fraud relationship networks include: 1. When extracting data features, use a single feature source from the data or text.

[0014] 2. When vectorizing text features, use the final output representation vector of the pre-trained language model as the feature vector representation. For example, take "pooler_output" from the final output of the Bidirectional Encoder Representations from Transformers (BERT) as the feature vector.

[0015] 3. A relational network is constructed using random sampling partitioning.

[0016] The aforementioned main methods for constructing anti-fraud relationship networks have several drawbacks. Based on a single feature source, they may not provide sufficiently comprehensive information to describe the relationships between objects, thus introducing bias and limitations; they may introduce information redundancy, i.e., multiple features provide similar or highly related information; and they may not be adaptable to diverse application scenarios. Using the final output of BERT as the feature vector may result in the inability to retain contextual information, leading to the loss of some semantic information. Using random sampling to construct the relationship network may lead to imbalanced data distribution and damage to the continuity and correlation structure of the data, thereby affecting the performance of the relationship network.

[0017] Currently, when auditing the fraud risk of individual insurance policies, the main approach is to use traditional relational database search combined with rule-based judgment. This involves auditors writing SQL statements that meet predetermined conditions based on their understanding of fraud risk, selecting policies from the existing relational database that match those conditions, and then manually judging each policy. The scope of policies extracted using this method is limited by the search logic written by the sales staff, resulting in limited accuracy. Furthermore, the need for manual judgment of each extracted policy is extremely time-consuming and labor-intensive. Therefore, the efficiency of this existing method needs improvement.

[0018] In the above application scenarios, existing anti-fraud solutions based on relational database filtering have the following problems: 1. The scope of audit coverage is easily limited. The number of problematic policies that business personnel can observe depends on the search logic they write. The number of policies that the search logic can cover is limited by the strength of the search logic. If the logic is too strong, the number of policies obtained will be less, which cannot meet the audit requirements. If the logic is too weak, it will cause false positives, obtaining a large number of policies that do not actually have fraud risks, thus affecting the effectiveness of the audit.

[0019] 2. The efficiency and accuracy of manual review are limited. In terms of efficiency, business personnel need to judge each item of the selected results, which greatly increases the tediousness of the work. At the same time, since the observed samples are selected through a given logic, the correctness of the logic depends on the business personnel's understanding of fraud risks. However, due to differences in experience and understanding, business personnel often cannot obtain a unified evaluation system, which affects the overall audit results. Furthermore, whether the given screening logic can cover most fraud cases is also questionable.

[0020] 3. Data samples obtained by filtering data through relational databases based on rules can only reflect simple, superficial relationships between individual policies, i.e. whether all policies meet the given filtering conditions. They cannot show the deep logical connections between policies, which is not conducive to judging the risk of group fraud.

[0021] This application aims to propose a method and apparatus for generating relationship graphs, in order to construct a relationship network that can more efficiently and accurately identify and judge the risk of policy fraud in audit business scenarios.

[0022] Compared to traditional relationship network construction methods, this application adopts a multi-dimensional feature combination analysis of policy information using text and data. After extracting features, the information is first clustered and then processed in batches. Combined with manual feedback and correction, an anti-fraud relationship network is constructed. This solves the problems of incomplete information coverage, unbalanced data distribution, and inability to adapt to diverse application scenarios in relationship networks. The constructed relationship network can more comprehensively and quickly find related target problematic policies, meeting the business needs of quickly locating gang fraud and tracing the source of problems in different scenarios.

[0023] Clustering generally refers to dividing all samples into different classes or groups according to a specific criterion, so that the similarity of data objects within the same group is maximized. In this embodiment, the clustering used utilizes feature vectors obtained by calculating combinations of different dimensions of features within a policy as input to the clustering algorithm, grouping each policy into several clusters. These clusters serve as the foundation for similar policy analysis, and subsequent relationship network construction is based on each cluster. Compared to constructing a relationship network directly from the original data, this approach improves the accuracy of the constructed network.

[0024] Regarding the relationship network, the relationship network referred to in this application is obtained by calculating the similarity of feature vectors obtained by combining different dimensional features of insurance policies and setting a certain threshold. It is used to determine whether there is a relationship between different insurance policies. The complexity of the relationship network is affected by the given threshold. When the threshold is low, there are more edges in the network, and vice versa.

[0025] The following is in conjunction with the appendix Figures 1 to 5 This application provides a detailed description of a method, apparatus, electronic device, and computer program product for generating a relational graph, through specific embodiments and application scenarios.

[0026] Figure 1 This application illustrates an embodiment of a method for generating a relationship graph. This method can be executed by an electronic device, which may include a server and / or terminal devices. In other words, the method can be executed by software or hardware installed on the server and / or terminal devices, and includes the following steps: Step 110: Obtain the textual and numerical information of the target object, and perform feature vectorization processing on the textual and numerical information to obtain the target feature vector.

[0027] The target objects typically contain data with multiple attributes, and the textual information includes unstructured or semi-structured data described in natural language.

[0028] This embodiment uses an insurance policy as the target object for specific explanation.

[0029] Textual information mainly includes address information associated with the insurance policy, such as the registered address, work address, and current residence address of the policyholder and the insured. This type of information is characterized by its rich semantic and spatial hierarchical relationships, but it cannot be directly used for numerical calculations.

[0030] Numerical information includes directly measurable structured data. Numerical information mainly involves policy costs and statistical indicators, such as total premiums paid, policy duration, insured's age, and historical claim count. While this type of information is already numerical, its dimensions and distribution often vary significantly, requiring standardization.

[0031] The feature vectorization process is used to provide a basis for subsequent analysis.

[0032] For feature vectorization of textual information, especially address information, pre-trained language models (e.g., BERT, RoBERTa) are used for deep semantic feature extraction. Specifically, the address text is first cleaned and standardized, then truncated using province, city, and district, separating the address into two parts: administrative region labels (e.g., "Zhejiang Province," "Hangzhou City") and the actual address content (e.g., "No. xx, Wensan Road, Xihu District"). The processed address content is then input into the pre-trained language model, extracting the output representations of multiple Transformer layers (e.g., all layers or several intermediate layers). These different layer outputs capture diverse information from low-level lexical and mid-level syntactic layers to high-level semantics. Finally, pooling or weighted concatenation is used to fuse these multi-layered representations into a unified, high-dimensional text feature vector. This vector not only includes the overall semantics of the address but also preserves its internal detailed structural information.

[0033] For feature vectorization of numerical information, the first step is preprocessing, which may include handling missing and outlier values. Then, standardization or normalization is performed to bring all numerical features to the same order of magnitude, preventing certain large numerical features from unreasonably dominating subsequent similarity calculations. All the selected numerical features after processing are arranged in a predefined order to form a one-dimensional array, i.e., the numerical feature vector.

[0034] After obtaining the text feature vector and the numerical feature vector, they are merged into a comprehensive target feature vector through vector concatenation. This target feature vector is the unique mathematical representation of the target object in the multi-dimensional feature space, and it integrates the deep semantics of its textual description and the quantitative attributes of numerical statistics, providing a data foundation for subsequent clustering analysis and relationship mining.

[0035] Step 120: Perform cluster analysis based on the target feature vector to obtain several clusters.

[0036] Cluster analysis divides a massive number of target objects (e.g., millions of insurance policies) into several subsets (i.e., clusters) with high internal similarity and significant differences from each other, based on the similarity of their multi-dimensional features (i.e., target feature vectors).

[0037] Several mature clustering algorithms can typically be used, such as K-Means clustering, DBSCAN density clustering, and hierarchical clustering. The choice of algorithm can be adjusted according to the data scale, distribution characteristics, and business requirements. Specifically, all target feature vectors are input into the selected clustering algorithm. The algorithm measures the similarity between objects by calculating the distance between vectors (e.g., Euclidean distance, cosine distance, etc.) and automatically divides the vectors into several clusters according to preset rules or parameters (e.g., the number of clusters, or density threshold, etc.), along with a label for the cluster to which each target object belongs. Each cluster contains several target feature vectors, representing a group of target objects that are close to each other in the feature space.

[0038] Step 130: Calculate the similarity between target feature vectors within each cluster, determine the connection relationship of the target feature vectors based on the similarity, and generate a relationship graph.

[0039] This step uses each cluster generated in step 120 as an independent unit to construct a relational graph.

[0040] Within each cluster, similarity calculations are performed. Specifically, cosine similarity is used to calculate the similarity between every two target feature vectors within that cluster. Cosine similarity focuses on the directional consistency of vectors in multidimensional space and is insensitive to the absolute magnitude of the values, making it suitable for measuring the similarity of high-dimensional fused feature vectors. After calculation, a similarity matrix is ​​obtained that describes the pairwise similarity between all objects within the cluster.

[0041] Based on business experience or experimental verification, a similarity threshold is preset. This threshold is used to control the sparsity and density, and the precision and breadth of the map.

[0042] The similarity matrix is ​​traversed. For any two target feature vectors, if their similarity value is greater than or equal to a preset threshold, a strong correlation is determined between the two target objects. In the graph, this translates to establishing an undirected edge between the nodes represented by these two target objects. If the similarity is below the threshold, no connection is established.

[0043] After performing the above steps for each cluster, a relational subgraph is obtained within that cluster. Each subgraph is an undirected graph consisting of nodes (target objects) and edges (relationships). Subgraphs from all clusters can be merged to form a complete global relational graph, depending on business requirements. The final output is a relational graph, which can contain a node set and an edge set. The node set represents all target objects, and the edge set represents the relationships between objects determined based on multi-dimensional feature similarity.

[0044] Step 140: Correct the relationship graph based on external feedback information.

[0045] External feedback includes supplementary knowledge derived from business execution processes or human judgment. For example, definitive judgments made by business auditors or investigators after reviewing specific policies or related groups; results of on-site verification of policies; lists of fraudulent policies obtained from business systems that have been definitively verified; and strong correlation rules recognized by business experts, even if not reflected in the raw data.

[0046] Specifically, the relationship graph can be modified based on the type of feedback information. Specifically, for confirmed connections not reflected in the graph, or newly discovered fraudulent entities, corresponding nodes and connecting edges can be added directly; for connections deemed erroneous, or closed / invalid entities, corresponding edges or nodes can be deleted; edges in the relationship graph can be assigned weights, with initial weights calculated from similarity and dynamically adjusted based on feedback information; business tags can be added to nodes, such as confirmed fraud, high risk, approved, etc., enriching the semantic information of the graph and aiding subsequent analysis.

[0047] In this embodiment, by fusing multi-dimensional information from textual and numerical sources to construct the target feature vector, the limitations of a single feature source can be overcome, resulting in a more comprehensive object representation. The approach of first clustering and then dividing to construct the graph maintains the local structure and continuity of the data distribution, avoiding structural damage caused by random sampling. This allows the generated relationship graph to more accurately reflect the deep, multi-dimensional connections between objects. Graph-based association retrieval reduces the workload of manual review, while clustering supports global or specific regional analysis, enabling auditing work to quickly focus on different scenario needs, thus expanding the coverage and flexibility of risk screening. Furthermore, introducing external feedback to correct the relationship graph allows it to continuously absorb business experience and adapt to changes in fraud methods, thereby improving the model's practicality and long-term effectiveness.

[0048] Figure 2 The diagram illustrates a flowchart of another method for generating a relationship graph according to an embodiment of this application. This method can be executed by an electronic device, which may include a server and / or terminal devices. In other words, the method can be executed by software or hardware installed on the server and / or terminal devices, and includes the following steps: Step 210: Based on step 110 of the above embodiment, obtain the textual and numerical information of the target object, and perform feature vectorization processing on the textual and numerical information to obtain the target feature vector. The method of this embodiment may further include the following specific steps: The textual information is input into a pre-trained language model to obtain the output representations of multiple network layers in the pre-trained language model. A textual feature vector is generated based on the output representations. The numerical information is standardized to generate a numerical feature vector. The textual feature vector and the numerical feature vector are concatenated to obtain the target feature vector.

[0049] Step 220: Perform cluster analysis based on the target feature vector to obtain several clusters.

[0050] Step 230: Calculate the similarity between target feature vectors within each cluster, determine the connection relationship of the target feature vectors based on the similarity, and generate a relationship graph.

[0051] Step 240: Correct the relationship graph based on external feedback information.

[0052] Steps 220-240 can be found above. Figure 1 The specific descriptions of steps 120-140 in the illustrated embodiment are provided, and the same technical effects can be achieved. To avoid repetition, they will not be repeated here.

[0053] In this embodiment, by fusing the deep semantics of textual information with the statistical quantification features of numerical information, the constructed target feature vector can more completely and three-dimensionally characterize the essential attributes of the target object from multiple dimensions, avoiding the one-sidedness and limitations of a single data source. Unifying unstructured text and structured numerical values ​​into mathematical vectors provides a standardized and computable data foundation for subsequent clustering, similarity calculations, etc. Compared to existing technologies that only use the final output of a pre-trained language model, this embodiment generates text feature vectors by extracting and fusing the output representations of multiple network layers in the model. This can simultaneously include low-level lexical information, mid-level syntactic combination information, and high-level overall semantic information, effectively overcoming the problem of contextual information loss caused by a single-layer representation, and improving the expressive power and discriminative power of text features. By standardizing numerical information, the unreasonable impact of differences in the units and value ranges of different features on the overall similarity calculation can be eliminated.

[0054] In yet another exemplary embodiment, based on step 210 of the above embodiment, textual and numerical information of the target object are obtained, and feature vectorization processing is performed on the textual and numerical information to obtain the target feature vector. The method of this embodiment may further include the following specific steps: When the text information includes address information, the address information is parsed in a structured manner to obtain administrative region labels and target address information. The target address information is then input into the pre-trained language model to participate in the generation of the text feature vector.

[0055] In this embodiment, address information is structured and parsed, splitting it into administrative region labels (such as province and city) and target address information (such as street and house number). This separates standardized, clearly categorized administrative regions from highly variable specific addresses, improving the clarity and targeting of information processing. Only the more discriminative target address information (i.e., specific addresses) is input into the model for deep semantic analysis, allowing the model to focus on identifying substantial, subtle differences between addresses, rather than being dominated by broad administrative affiliations. This enhances the sensitivity of features to differences in actual geographical locations. The separated administrative region labels provide a basis for filtering and divide-and-conquer in subsequent cluster analysis. Auditors can flexibly choose to conduct a nationwide global analysis or focus on in-depth audits of specific provinces or cities, enhancing the method's practicality and business coverage. In local audit scenarios, data can be pre-filtered based on administrative labels, reducing unnecessary computation and focusing the analysis on risk patterns within the target area, improving audit efficiency and targeting.

[0056] In yet another exemplary embodiment, this embodiment uses an insurance policy as the target object for specific illustration. The obtained policy information is combined and concatenated using features from multiple dimensions, categorizing the target object's information into two main types: textual information and numerical information. Based on the daily reference definitions used by business personnel, textual information mainly pertains to the policyholder, the insured's registered address, work address, and current residence address; numerical information includes a set of information related to policy fees, such as accumulated premiums. The obtained numerical feature vector and textual feature vector are then concatenated horizontally to obtain the final policy feature vector used.

[0057] For text-based address information, a pre-trained language model is used to vectorize its features. Using the outputs of each transformer module in the BERT model, several output representations are obtained. These representations are averaged and then transformed into a final feature representation, resulting in a vector representation that contains multi-scale compressed information to improve its interpretability.

[0058] It should be noted that the address processing adopts the province, city, and district truncation method. Only the actual address other than the province and city is used as the input for feature calculation, while the truncated province and city fields are used as a filtering condition for the subsequent clustering algorithm to distinguish the implementation effect.

[0059] In yet another exemplary embodiment, based on step 120 of the above embodiment, cluster analysis is performed based on the target feature vector to obtain several clusters. The method of this embodiment may further include the following specific steps: Based on the administrative region label, a subset of target objects that meet the set conditions is selected from the target objects; the target feature vectors corresponding to each target object in the subset of target objects are clustered, and the clustering includes global clustering and / or local clustering.

[0060] In this embodiment, filtering based on administrative region labels (such as province and city) supports both global and localized special audits, directly meeting the business needs of audit departments at different levels. In localized audit scenarios, all computing resources and audit attention can be concentrated on specific high-risk or key areas, avoiding a full data search and enabling precise allocation of risk assessment resources. Because the input data subset is geographically homogeneous, eliminating interference from heterogeneous data from other regions, the clustering algorithm can capture the unique risk patterns and correlation characteristics within that region, resulting in clusters with higher cohesion and stronger business interpretability. Filtering also reduces the amount of data involved in clustering calculations, lowering computational complexity and time overhead.

[0061] In yet another exemplary embodiment, this embodiment specifically describes the target object as an insurance policy. Using the target feature vector obtained in the above embodiment, a clustering algorithm is executed to group the various samples into several clusters, which serve as a set for constructing a relationship graph. Using the administrative region labels obtained in the above implementation, such as province and city labels obtained by address splitting, full-sample clustering and clustering only for samples from specific provinces or cities can be performed to adapt to the requirements of different audit scenarios, i.e., whether it is a global audit of the entire sample or a specialized audit of only a provincial or municipal branch.

[0062] In yet another exemplary embodiment, based on step 130 of the above embodiment, the similarity between target feature vectors within each cluster is calculated, the connection relationship of the target feature vectors is determined according to the similarity, and a relationship graph is generated. The method of this embodiment may further include the following specific steps: Calculate the cosine similarity between every two target feature vectors within each cluster; for target feature vector pairs whose cosine similarity exceeds a preset similarity threshold, set connection edges between the nodes corresponding to the target feature vector pairs.

[0063] In this embodiment, cosine similarity is used to measure the directional consistency of high-dimensional feature vectors, ignoring the influence of their absolute numerical magnitude. This quantifies the overall similarity between target objects in the multi-dimensional feature space, providing a scientific and calculable basis for determining whether potential associations exist between objects. By introducing a preset similarity threshold, continuous similarity values ​​are transformed into explicit binary connection decisions, i.e., edge-based or edgeless, resulting in a clear graph structure that facilitates subsequent graph algorithm processing and manual interpretation. Connections are established only between vector pairs with cosine similarity exceeding the threshold, ensuring that each edge in the graph represents a quantitatively verified, strong feature similarity. This significantly reduces the false alarm rate, giving the graph-presented association network a high degree of confidence and supporting subsequent auditing work.

[0064] In yet another exemplary embodiment, this embodiment uses an insurance policy as the target object for specific illustration. For each cluster obtained from clustering, the cosine similarity of the feature vectors of samples within the cluster is calculated to determine whether there is an edge connecting two samples.

[0065] Following the aforementioned steps, all policies have been clustered into several clusters based on their target feature vectors using a clustering algorithm. These target feature vectors have been integrated with processed numerical features such as address semantics and premiums. Policies within each cluster exhibit high overall similarity in the feature space and can be considered a potentially related group. For example, a cluster might contain 50 policies whose feature vectors are spatially close, possibly due to shared regional address characteristics and a pattern of high coverage and short-term premium payments.

[0066] Clustering does not perform pairwise comparisons across all policies; instead, it performs the following operations independently for each cluster. First, it calculates similarity: for all policies within a cluster, it calculates the cosine similarity between the target feature vectors of each pair of policies. For example, among these 50 policies, the cosine similarity between policy A and policy B is 0.92, and the similarity with policy C is 0.45. Next, it determines a threshold and establishes an edge: a preset similarity threshold (e.g., 0.85) is used. If the cosine similarity between two policies is greater than or equal to this threshold, a strong association is determined between the two policies, and an undirected edge is established between their corresponding nodes in the relationship graph. If the similarity is below the threshold, no connection is established.

[0067] After performing the above operations on each cluster, each cluster generates an internal relational subgraph. Merging the subgraphs of all clusters yields the complete global policy relationship graph. In this graph, nodes represent each independent policy, and edges represent two policies that the algorithm determines to have a strong relationship.

[0068] In yet another exemplary embodiment, based on step 140 of the above embodiment, the relationship graph is corrected according to external feedback information. The method of this embodiment may further include the following specific steps: The association information and / or manual verification information of the target object are obtained from the business side and used as the external feedback information.

[0069] In this embodiment, by introducing correlation information from the business side and manual verification information, the domain knowledge, on-site experience, and real-time judgment of human experts are injected into the relationship graph generation process. This directly corrects potential biases, misjudgments, or omissions that may arise from purely data-driven models, performing real-time calibration and completion of the graph to ensure it reflects the true business situation. Manual verification can directly eliminate false correlations generated by the algorithm, while the correlation information from the business side can supplement strong correlations that the algorithm fails to detect, thereby improving the confidence of each connection edge in the graph. The integration of business-side correlation information into the graph allows it to not only display data similarities but also provide business-related correlations, enhancing the practicality of the relationship graph.

[0070] In yet another exemplary embodiment, this embodiment uses an insurance policy as the target object for specific explanation. The generated relationship graph already covers policy-related text and numerical information. Considering that the policy may involve on-site investigations by agents, the relationship graph can be modified based on the information returned by agents regarding the policyholder, the insured's family, workplace, etc., from the on-site investigations.

[0071] As mentioned above, the association graph obtained from the above steps is an undirected graph structure, which can be traversed and searched using a fast graph search algorithm. When a certain insurance policy sample is involved in fraud, the generated graph structure can be used to quickly locate the related insurance policies for key audits to prevent the occurrence of gang fraud risks.

[0072] In yet another exemplary embodiment, the textual information includes address information, and the numerical information includes quantitative indicator information associated with the target object.

[0073] In this embodiment, by concretizing textual information into address information and numerical information into quantitative indicator information, the two most critical and discriminative types of data in anti-fraud analysis can be anchored. Address information is used to identify core clues for false identities and related groups, while quantitative indicator information is used to measure insurance behavior and assess risk exposure, thereby improving search efficiency in multi-source heterogeneous data.

[0074] Corresponding to the relationship graph generation method provided in the above embodiments, based on the same technical concept, this application also provides a relationship graph generation apparatus. See [link to related documentation] Figure 3 The device 400 includes a feature generation module 410, a clustering processing module 420, a map generation module 430, and a feedback correction module 440.

[0075] The feature generation module 410 is used to acquire textual and numerical information of the target object, perform feature vectorization processing on the textual and numerical information to obtain target feature vectors; the clustering processing module 420 is used to perform clustering analysis based on the target feature vectors to obtain several clusters; the graph generation module 430 is used to calculate the similarity between target feature vectors within each cluster, determine the connection relationship of the target feature vectors based on the similarity, and generate a relationship graph; and the feedback correction module 440 corrects the relationship graph based on external feedback information.

[0076] It should be noted that the relation graph generation apparatus and the relation graph generation method provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned relation graph generation method, and the repeated parts will not be described again.

[0077] The method and apparatus for generating relationship graphs provided in this application's embodiments are described in [reference needed]. Figure 4This process generates a complete policy relationship graph that meets audit requirements. This graph allows auditors to overcome the limitations of relational databases and quickly locate policies associated with problematic policies using graph search algorithms, further preventing the risk of organized fraud. The generation of multi-dimensional feature vectors for policies, especially in the text-to-vector conversion process, employs a method of concatenating feature matrices generated by various transformer modules in the BERT pre-trained model. This considers feature vectors at multiple scales, increasing the information content of the final feature vectors and further improving the accuracy of similarity calculations. Corresponding to the relationship graph generation method provided in the above embodiments, based on the same technical concept, this application also provides an electronic device for executing the above method. Figure 5 To illustrate the structure of an electronic device according to various embodiments of this application, as shown in the following diagrams... Figure 5 As shown. Electronic device 500 can vary considerably due to differences in configuration or performance, and may include one or more processors 510 and memory 520. Memory 520 may store one or more application programs or data. Memory 520 may be temporary or persistent storage. The application programs stored in memory 520 may include one or more modules (not shown), each module may include a series of computer-executable instructions for the electronic device. Furthermore, processor 510 may be configured to communicate with memory 520 and execute the series of computer-executable instructions stored in memory 520 on the electronic device.

[0078] This application also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer, implement the steps of the relationship graph generation method described above.

[0079] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0080] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems, devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0082] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0083] In a typical configuration, an electronic device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0084] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0085] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0086] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0087] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0089] It should be understood that the training and prediction processes of the artificial intelligence (AI) models involved in the various embodiments of this specification all adhere to multiple legal and compliant principles, including legal data sources, compliant data content, compliant data governance, compliant training objectives and schemes, compliant training processes, compliant training environments and tools, and compliant ethical verification of training results, and comply with the requirements of Article 5 of the Patent Law. Among them: Data source legitimacy: All datasets used for AI model training were obtained through legal means, covering three categories: publicly authorized data, data authorized by partners, and self-collected compliant data. Publicly authorized data comes from compliant data sources following open-source licenses such as Apache 2.0, with complete copyright attribution and authorization scope clearly marked, and no unauthorized open-source code or data reuse. Data authorized by partners has been subject to formal data usage agreements, clearly defining the scope, duration, and confidentiality obligations, and possessing a complete authorization chain. For self-collected data involving personal information, strict informed consent procedures have been followed, and anonymization processes (including but not limited to field masking, feature anonymization, and differential privacy technology applications) have been implemented to remove personally identifiable information, fully complying with the requirements of relevant laws and regulations such as the "Interim Measures for the Administration of Generative Artificial Intelligence Services" and the "Personal Information Protection Law."

[0090] Data content compliance: The AI ​​model's dataset undergoes multiple screenings and cleaning processes to remove all content that may violate social morality or harm public interests. It contains no obscene, pornographic, violent, discriminatory, or information that endangers national or public safety, nor does it involve the illegal acquisition or use of genetic resources. For data in sensitive fields (such as healthcare and finance), an additional privacy-preserving computation module (including federated learning and secure multi-party computation technologies) ensures that the data is "usable but not visible," avoiding compliance risks during the original data transmission process and ensuring that the data application scenarios and uses comply with public order and good morals and industry regulatory requirements.

[0091] Data governance norms: A complete data traceability system is established during the AI ​​model training process to automatically record the source, collection time, annotation process, cleaning rules, and permission allocation of training data, generating traceable compliance reports to ensure that the data is verifiable throughout its entire lifecycle. The dataset annotation process for AI models is completed by a professional human R&D team, clearly defining the proportion of human creative contributions and avoiding reliance on AI-generated data that has not undergone substantial human modification, thus meeting the examination requirements for "human main contributions" in AI patent applications.

[0092] Training objectives and plans are compliant: The training objective of the AI ​​model focuses on [specific technical scenarios that can be supplemented, such as intelligent driving decision optimization, multimodal information interaction, etc., and replaced based on specific content]. The training scheme and the final output results do not violate any mandatory provisions of laws and administrative regulations, do not harm the public interest or the legitimate rights and interests of others, and do not pose any potential risks of being used for illegal activities, infringing on privacy, or disrupting public safety. The model strictly adheres to the ethical principle of "intelligent for good".

[0093] Training process compliance: A closed-loop training framework is adopted to ensure compliance and controllability of the training process. The specific process is as follows: First, training samples are obtained through compliant data sources. After the aforementioned data cleaning and desensitization, they are input into the neural network model to generate preliminary training results. Second, an expert system is introduced to verify the preliminary results. Based on preset rules and human expert experience, the feasibility of the results is evaluated, and outputs that may pose ethical risks or compliance hazards are corrected (such as removing decision-making logic that violates public order and good morals, and adjusting model parameters that do not comply with safety regulations). Finally, the loss function weights are dynamically optimized based on expert system feedback to strengthen the model's learning of compliant results, avoid overfitting errors or non-compliant labels, and form a closed-loop control of "data input - model training - expert verification - parameter optimization - result feedback" to ensure that the entire training process complies with A5 ethical review requirements.

[0094] Training environment and tool compliance: AI model training is implemented using nationally licensed chips and a compliant training platform. All open-source frameworks and components used in the training process have obtained their corresponding licenses, and copyright statements and patent citation information are fully retained, with no instances of infringement or reuse. The training environment is built using virtual devices (containers / virtual machines) with fixed random seeds and initial parameter configurations to ensure the reproducibility of the training process. Furthermore, through access control and operation log recording, risks such as data leakage and parameter tampering during training are prevented, ensuring the security and compliance of the training process.

[0095] Training results ethical verification compliance: After the model is trained, it undergoes additional third-party ethical compliance assessment and algorithm filing review to verify that the model output does not violate social morality or harm public interests. For potentially sensitive scenarios (such as public services and intelligent decision-making), a special result verification mechanism is established to ensure that the model always complies with Article 5 of the Patent Law and relevant laws and regulations in practical applications.

[0096] In summary, the data and training process used in the AI ​​model of this specification strictly comply with the relevant provisions of Article 5 of the Patent Law and the Patent Examination Guidelines (2023 Edition), and there are no violations of laws, social ethics, public interests, or illegal use of genetic resources. It fully meets the compliance requirements for patent authorization.

Claims

1. A method for generating a relational graph, characterized in that, The method includes the following steps: Obtain textual and numerical information of the target object, and perform feature vectorization processing on the textual and numerical information to obtain the target feature vector; Cluster analysis is performed based on the target feature vectors to obtain several clusters; Calculate the similarity between target feature vectors within each cluster, determine the connection relationship of the target feature vectors based on the similarity, and generate a relationship graph; The relationship graph is revised based on external feedback.

2. The method according to claim 1, characterized in that, The process of obtaining textual and numerical information of the target object, and performing feature vectorization on the textual and numerical information to obtain the target feature vector, includes the following steps: The text information is input into a pre-trained language model to obtain the output representation of multiple network layers in the pre-trained language model, and a text feature vector is generated based on the output representation. The numerical information is standardized to generate numerical feature vectors; The target feature vector is obtained by concatenating the text feature vector and the numerical feature vector.

3. The method according to claim 2, characterized in that, The step of obtaining the textual and numerical information of the target object, and performing feature vectorization processing on the textual and numerical information to obtain the target feature vector, further includes the following steps: When the text information includes address information, the address information is parsed in a structured manner to obtain administrative region labels and target address information. The target address information is then input into the pre-trained language model to participate in the generation of the text feature vector.

4. The method according to claim 3, characterized in that, The clustering analysis based on the target feature vector to obtain several clusters includes the following steps: Based on the administrative region label, a subset of target objects that meet the set conditions is selected from the target objects; Clustering is performed on the target feature vectors corresponding to each target object in the subset of target objects, and the clustering includes global clustering and / or local clustering.

5. The method according to claim 1, characterized in that, The process of calculating the similarity between target feature vectors within each cluster, determining the connectivity of the target feature vectors based on the similarity, and generating a relationship graph includes the following steps: Calculate the cosine similarity between every two target feature vectors within each cluster; For target feature vector pairs whose cosine similarity exceeds a preset similarity threshold, connection edges are set between the nodes corresponding to the target feature vector pairs.

6. The method according to claim 1, characterized in that, The step of correcting the relationship graph based on external feedback information includes the following steps: The association information and / or manual verification information of the target object are obtained from the business side and used as the external feedback information.

7. The method according to any one of claims 1 to 6, characterized in that, The textual information includes address information, and the numerical information includes quantitative indicator information associated with the target object.

8. A device for generating a relational graph, characterized in that, The device includes the following: The feature generation module is used to acquire textual and numerical information of the target object, and to perform feature vectorization processing on the textual and numerical information to obtain the target feature vector. The clustering processing module is used to perform clustering analysis based on the target feature vector to obtain several clusters; The graph generation module is used to calculate the similarity between target feature vectors within each cluster, determine the connection relationship of the target feature vectors based on the similarity, and generate a relationship graph. as well as, The feedback correction module corrects the relationship graph based on external feedback information.

9. An electronic device, characterized in that, It includes a processor, a memory, a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions that, when executed by a computer, implement the steps of the method as described in any one of claims 1 to 7.

Images