Customer archive knowledge graph construction method and system based on large model driving

Abstract

The invention relates to the field of knowledge graph construction, in particular to a customer archive knowledge graph construction method and system based on large model driving. The method comprises the following steps: encoding customer archives by using a large model technology, and fusing encoding results to generate entity twin vectors; constructing a knowledge graph by using structured data in the customer archive and the entity twin vector; the customer archive comprises an interaction log, and calculating an edge relation weight of the knowledge graph by using the interaction log; and adjusting the knowledge graph based on the edge relation weight. The method solves the problem that in the prior art, due to the fact that a traditional knowledge graph is difficult to accurately capture entity dynamic characteristics and implicit association, the constructed knowledge graph is difficult to meet the dynamic business requirements of real-time risk control, precision marketing and the like. And the interpretability and the reasoning ability of the knowledge graph are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph construction, and specifically to a method and system for constructing a customer profile knowledge graph driven by a large model. Background Art

[0002] As a structured knowledge representation method, knowledge graphs have broad application prospects in areas such as customer profile management and business relationship analysis. Traditional customer profile management is mostly based on relational databases, which can only store static entity attributes and explicit associations, but have difficulty capturing dynamic entity features and implicit associations. This results in insufficient dynamic and reasoning capabilities of knowledge graphs.

[0003] With the development of big data technology, customer profile data presents multimodal, temporal and high-dimensional characteristics. Traditional graph construction methods based on rule engines or shallow neural networks face many challenges, such as inefficient multimodal feature fusion, and the difficulty of effectively integrating text semantic ambiguity, behavioral temporal dependency, and image visual features through traditional feature engineering. This leads to a single dimension of entity representation, an inability to fully characterize customer dynamic characteristics, limited implicit relationship mining capabilities, reliance on predefined rules or statistical associations, difficulty in capturing deep cross-modal semantic associations and generating potential relationship hypotheses, and inability to update entity features and relationship weights in real time, making it difficult to support real-time risk control, precision marketing and other business needs.

[0004] Therefore, there is an urgent need for a knowledge graph construction method that deeply explores the relationships between entities so that the knowledge graph can better characterize multi-dimensional entities. Summary of the Invention

[0005] In response to the defects in the existing technology, the present invention provides a method and system for constructing a customer profile knowledge graph based on a large model drive, which solves the problem that traditional knowledge graphs in the existing technology are difficult to accurately capture the dynamic characteristics and implicit associations of entities, resulting in the constructed knowledge graph being difficult to meet business needs such as real-time risk control and precision marketing.

[0006] In order to achieve the above-mentioned objectives, the present invention provides a method for constructing a customer profile knowledge graph driven by a big model, comprising encoding customer profiles using big model technology, and fusing the encoding results to generate entity twin vectors; constructing a knowledge graph using the structured data in the customer profiles and the entity twin vectors; the customer profiles include interaction logs, and the edge relationship weights of the knowledge graph are calculated using the interaction logs; and adjusting the knowledge graph based on the edge relationship weights.

[0007] This invention uses large-scale model technology to perform multimodal encoding and fusion of customer profiles, generating entity twin vectors that evolve over time. This breaks through the static and single limitations of entity features in traditional methods and can dynamically capture real-time changes in customers in dimensions such as text semantics, behavioral patterns, and visual attributes. By constructing a knowledge graph using customer profiles and twin vectors, it achieves a deep binding of explicit associations (such as transaction records and contractual relationships) with entity features, while providing a multi-dimensional feature foundation for implicit relationship mining. By calculating edge relationship weights with the help of interaction logs and incorporating the time dimension into graph modeling, the strength of associations between entities can be quantified to reflect temporal patterns such as interaction frequency and time intervals, improving the graph's ability to depict the dynamic evolution of business.

[0008] Optionally, the customer file includes text data, behavior data and image data, and the use of large model technology to encode the customer file and fuse the encoding results to generate an entity twin vector includes: using large model technology to extract semantic features of the text data, extract behavioral features of the behavior data, and extract visual features of the image data to obtain semantic vectors, behavior vectors and visual vectors; converting the semantic vectors, behavior vectors and visual vectors into semantic description texts, behavior description texts and visual description texts respectively; and using a dynamic gating mechanism to fuse the semantic vectors, behavior vectors, visual vectors, semantic description texts, behavior description texts and visual description texts to obtain entity twin vectors.

[0009] This invention utilizes a large model to extract specialized features specific to different modal data characteristics, achieving deep decoupling and precise characterization of multimodal features, avoiding the limitations of traditional feature engineering. It converts behavioral and visual vectors into natural language descriptions, eliminating the multimodal semantic gap through the semantic generation capabilities of the large model, and laying a unified semantic foundation for cross-modal fusion. It also employs a dynamic gating mechanism to fuse multiple modalities, ensuring that entity twin vectors reflect changes in the entity's core features in real time, thereby improving the accuracy of entity twin vectors.

[0010] Optionally, the entity twin vector satisfies the following formula: in, Over time Updated entity twin vector, is the Sigmoid function, For pre-trained generative large language models, For the Mode in time Description text of For the Mode in time The eigenvector of .

[0011] The present invention processes multimodal description texts through a generative large language model, and combines the Sigmoid function to dynamically weight and fuse the feature vectors of each modality to achieve entity representation updates that evolve over time. It not only uses the semantic understanding ability of the generative large language model to mine the value of multimodal data, but also uses the dynamic gating mechanism to accurately adapt to the importance of each modality at different times, further improving the accuracy of entity twin vectors.

[0012] Optionally, the construction of a knowledge graph using the structured data in the customer file and the entity twin vector includes: extracting business rules based on the business needs of the target field, and obtaining entities based on the structured data in the customer file and the business rules; extracting direct associations between the entities using the business data in the structured data, and associating the direct associations with the entity twin vectors to obtain explicit relationships; analyzing the entity twin vectors corresponding to the entities through a generative big model to obtain potential relationships; and constructing a knowledge graph based on the entities, the explicit relationships and the potential relationships.

[0013] The present invention uses the business needs of the target field as an anchor, combines the structured data of customer files with business rules to accurately extract entities, and ensures that the entities fit the business logic. Business data is used to mine direct associations between entities, and then the entity twin vectors are associated to form explicit relationships, so that the entity associations can retain the direct connection between business scenarios and integrate the dynamic characteristics of entities, making the relationship expression richer and more three-dimensional. With the help of generative large models to analyze the dynamic characteristics of entity twin vectors to mine potential relationships, it breaks through the limitations of traditional reliance on explicit business data, and can discover subtle relationships such as customer behavior trend associations and cross-scenario implicit interactions, which expand the depth and breadth of relationships in the knowledge graph. The knowledge graph constructed by the collaboration of entities, explicit relationships and potential relationships not only consolidates the basic business associations, but also extends the mining of implicit value to improve the knowledge-driven ability of the knowledge graph.

[0014] Optionally, obtaining entities based on the structured data of the customer file and the business rules includes: vectorizing the structured data of the customer file to obtain a structured data vector; weightedly fusing the structured data vector and the entity twin vector to construct an entity feature matrix; analyzing the entity feature matrix using a generative big model to generate a set of candidate entities; introducing a rule engine, using the rule engine to load the business rules and screening and standardizing the candidate entity set to obtain multiple entities.

[0015] Optionally, the structured data of the customer file includes numerical structure data and categorical structure data, and the vectorization processing of the structured data of the customer file to obtain the structure data vector includes: normalizing the numerical structure data to generate a numerical feature vector; pre-training embedding coding the categorical structure data to generate a categorical feature vector; and splicing the numerical feature vector and the categorical feature vector to obtain the structure data vector.

[0016] This method vectorizes structured data and combines it with entity twin vectors to construct a feature matrix, fully integrating the advantages of multi-source data to enrich and comprehensively enrich entity features. Generative large-scale model analysis matrices are used to generate candidate entities, mining potential entity information and breaking through the limitations of traditional methods. A rule engine is then introduced to screen and standardize candidate entities based on business rules, ensuring that entities meet business requirements while standardizing entity forms and improving entity quality and consistency.

[0017] Optionally, the analyzing the entity twin vector corresponding to the entity through the generative big model to obtain the potential relationship includes: using the generative big model to analyze the entity twin vector corresponding to the entity to obtain the potential hypothesis relationship and confidence; constructing a predefined rule base, and using the confidence based on the predefined rule base to identify the potential hypothesis relationship to obtain the potential relationship.

[0018] This paper, through a generative large-scale model, deeply analyzes entity twin vectors, capturing cross-modal and temporal semantic associations that are difficult to detect with traditional methods. It generates potential hypothetical relationships and quantifies their confidence, injecting new knowledge discovery capabilities into the graph. The introduction of a predefined rule library effectively filters out low-quality hypothetical relationships or those that do not conform to domain knowledge through business logic constraints, while also being consistent with business practices, and enhancing the scientific nature and accuracy of potential relationship mining.

[0019] Optionally, the method of calculating the edge relationship weight of the knowledge graph using the interaction log includes: adding the time information in the interaction log to the knowledge graph to form a time series knowledge graph; extracting the time series data of the edge relationship in the time series knowledge graph to form a time series vector; constructing a weight calculation model based on a time series graph neural network, and obtaining the edge relationship weight using the weight calculation model according to the time series vector and the semantic vector.

[0020] By introducing time information from interaction logs, this paper upgrades the knowledge graph to a time-series knowledge graph, enabling dynamic temporal modeling of edge relationship weights. By extracting the time-series data of edge relationships and converting it into vectors, and then integrating this vector into a time-series graph neural network model, it effectively captures the temporal dependencies of entity associations and the linkage of semantic features. The introduction of the time-series graph neural network enables the model to learn how relationship strength evolves over time, ultimately improving the accuracy of weight calculations.

[0021] Optionally, the adjusting of the knowledge graph based on the edge relationship weight includes: setting a weight threshold, comparing the weight threshold with the edge relationship weight; and updating or deleting the edge corresponding to the edge relationship weight in the knowledge graph according to the result of the comparison.

[0022] By setting weight thresholds and comparing edge relationship weights, this method dynamically filters out invalid or weakly associated edges while retaining strong relationships. This mechanism ensures that the knowledge graph remains focused on key business relationships, avoids redundant information interference, and improves the simplicity and readability of the knowledge graph.

[0023] Another aspect of the present invention provides a customer profile knowledge graph construction system driven by a large model, comprising: a processor, an input device, an output device and a memory, wherein the processor, the input device, the output device and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute a customer profile knowledge graph construction method driven by a large model as described in any one of the previous aspects of the present invention.

[0024] The present invention is a customer profile knowledge graph construction system based on large model drive, which has a compact structure, stable performance, high integration and simple structure. It can stably execute the customer profile knowledge graph construction method based on large model drive provided in the previous aspect of the present invention, further improving the overall applicability and practical application ability of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flowchart of a system for building a customer profile knowledge graph driven by a large model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a customer profile knowledge graph construction system based on a large model drive according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] Specific embodiments of the present invention will be described in detail below. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the present invention. In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention. However, it will be apparent to one of ordinary skill in the art that these specific details are not necessarily required to practice the present invention. In other instances, well-known circuits, software, or methods are not specifically described to avoid obscuring the present invention.

[0027] Throughout this specification, references to "one embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with the embodiment or example is included in at least one embodiment of the present invention. Therefore, appearances of the phrases "in one embodiment," "in an embodiment," "an example," or "an example" in various places throughout this specification are not necessarily all referring to the same embodiment or example. Furthermore, the particular features, structures, or characteristics may be combined in any suitable combinations and / or subcombinations in one or more embodiments or examples. Furthermore, those of ordinary skill in the art will appreciate that the figures provided herein are for illustrative purposes only and are not necessarily drawn to scale.

[0028] In an alternative embodiment, see Figure 1 ,like Figure 1 The method for constructing a customer profile knowledge graph based on a large model drive includes the following steps: Step S1: Encode the customer file using the big model technology, and fuse the encoding results to generate entity twin vectors.

[0029] The steps of encoding customer files using large model technology and fusing the encoding results to generate entity twin vectors specifically include the following: Step S101 , using a large model technology to extract semantic features from the text data, extract behavioral features from the behavioral data, and extract visual features from the image data, to obtain a semantic vector, a behavioral vector, and a visual vector.

[0030] In this embodiment, the customer profile includes text data, behavioral data, and image data, with the customer profile representing archival data for a predetermined period (e.g., two years). The structured data in the customer profile typically includes a unique identifier (ID). This unique identifier can be used to pre-categorize the text, behavioral, and image data in the customer profile. Therefore, the semantic, behavioral, and visual vectors obtained after feature extraction can be associated using the unique identifier.

[0031] Large model technology is used to extract semantic features from text data to obtain semantic vectors. First, the text data is preprocessed and converted into a token sequence through a word segmenter. Start and end symbols are added, and truncation or padding is performed to adapt to the model input length. Then, a pre-trained generative large language model such as LLaMA-2 or GPT-4 is selected. The pre-processed token sequence is input into the generative large language model. The contextual representation of each token is calculated through the Transformer layer of the generative large language model. The vector of the specific token is then extracted from the last hidden state layer of the model as the semantic representation of the text. Finally, the semantic vector is mapped to a unified dimensional space through a linear projection layer to form the final semantic vector. In this process, pre-trained models of different sizes can be selected according to the task requirements, and the semantic extraction effect of specific fields can be further optimized through fine-tuning.

[0032] Big model technology is used to extract behavioral features from behavioral data to obtain behavioral vectors. First, time series feature engineering is performed on behavioral data (such as user clicks, transactions, browsing and other sequence data) to convert them into standardized time series containing timestamps, behavior types, and numerical indicators, and basic features such as time intervals, behavior frequencies, and trend changes are extracted. Then, a pre-trained time series big model is used to input the time series feature sequence into the model. The time dependency, periodic patterns, and abnormal fluctuation characteristics of the behavioral data are captured through the model's attention mechanism or recurrent neural network layer. The feature vectors of each time step output by the model are aggregated into fixed-dimensional behavior vectors through global average pooling, maximum pooling, or self-attention weighting. Finally, the dimension of the behavior vector is adjusted through a linear transformation layer, thereby completing the feature extraction process from behavioral data to behavior vectors.

[0033] Large model technology is used to extract visual features from image data to obtain visual vectors. First, the image data is preprocessed, including scaling to a standard size, normalizing pixel values, random cropping or horizontal flipping to enhance data diversity, and targeted processing (such as contrast adjustment) based on the image type. Then, a pre-trained large visual model (such as CLIP, Swin Transformer, or ResNet-50 convolutional neural network) is selected, and the preprocessed image is input into the model. The convolutional layer of the model automatically extracts local features (such as texture, edges) and global semantic information (such as object categories and spatial relationships) of the image. Then, the feature map is obtained from the last convolutional layer or global pooling layer of the model, and it is compressed into a fixed-length feature vector through average pooling or RoI pooling. Finally, the visual vector is mapped to a unified dimensional space through a linear projection layer to form the final visual vector.

[0034] The creation time of text data, the interaction time of behavioral data, and the shooting time of image data are uniformly converted into Unix timestamps with an accuracy of seconds. For data with a time deviation exceeding a preset threshold (such as 1 hour), linear interpolation is used to correct it (for example, if the shooting time of an image differs from the corresponding transaction time by 2 hours, the feature vector of the intermediate moment is generated by interpolation in chronological order) to achieve cross-modal time alignment.

[0035] Step S102 : Convert the semantic vector, the behavioral vector, and the visual vector into a semantic description text, a behavioral description text, and a visual description text, respectively.

[0036] In this embodiment, for semantic vectors, a Seq2Seq model fine-tuned with industry texts (such as contracts and interaction logs) is used, combined with prompt word engineering (such as "extracting contract terms, interaction demands, and emotional tendencies"). The business themes, key entities, terms and demands, and other features implied in the semantic vectors are decoded into structured natural language descriptions (for example, "In June 2025, Customer A and Supplier B signed a three-year framework agreement with a 45-day payment period"). Beam search is used to optimize text fluency, and the time anchor and subject identifier are unified with the behavioral and visual description texts to achieve alignment of multimodal features in the time and business dimensions. For behavioral vectors, the pre-trained large-scale time series model T5-TimeSeries is used to decode behavioral features such as time intervals, behavior frequency, and trend changes in the behavioral vectors. Natural language descriptions are then generated through the T5-TimeSeries Transformer layer. For example, a behavioral vector describing "high-frequency clicks + fluctuating spending" is converted into a textual description describing "the user clicked on the product detail page more than five times per day in the past 30 days, and their spending increased by 40% compared to the previous month." For visual vectors, pre-trained visual language models such as BLIP and CLIP, or the image description generation model Show and Tell, are used to map image semantic features such as color, texture, and object category captured in the vectors into text. For example, a visual vector describing "night scene of high-rise buildings + traffic lights" is converted into a textual description describing "a bustling city at night, with tall buildings and dense traffic." This conversion process uses prompt engineering to guide the model in generating structured descriptions (such as short sentences containing time, location, and behavior type). Decoding strategies such as beam search are then used to optimize text fluency and information accuracy.

[0037] Step S103: A dynamic gating mechanism is used to fuse the semantic vector, the behavioral vector, the visual vector, the semantic description text, the behavioral description text, and the visual description text to obtain an entity twin vector.

[0038] In this embodiment, a pre-trained generative large language model such as LLaMA-2 is used to perform semantic analysis on the descriptive text of each modality, generate gating weights corresponding to the modality, and normalize the weights to dynamic gating coefficients between 0 and 1 through the Sigmoid function to reflect the importance of each modality at the current time. The normalized gating coefficients are weighted element-by-element with the feature vectors of the corresponding modalities, and then the weighted results of all modalities are summed to obtain the entity twin vector that evolves over time.

[0039] The generative large language model learns the temporal importance of each modal feature through the training and inference stages, uses historical multimodal data (with timestamps) to build a training set, and combines the modal description texts of the same entity at different time points. As input, we label the business importance of each modality at the corresponding time (for example, the semantic vector weight of "quarterly financial report text" in the financial scenario should be higher than that of daily behavior data), and optimize the model's ability to represent the importance of time series through supervised learning. Input , the generative large language model outputs a weight vector with the same dimension as the modal feature vector, and the gating coefficient is obtained after normalization by the Sigmoid function , when a mode is at time When new data (such as sudden transaction behavior) triggers the preset incremental threshold (such as the data volume exceeds the historical average by 20%), the gate weight will be updated in real time to ensure Prioritize the integration of key features.

[0040] The entity twin vector satisfies the following formula: in, Over time Updated entity twin vector, is the Sigmoid function, For pre-trained generative large language models, For the Mode in time Description text of For the Mode in time The eigenvector of .

[0041] In this embodiment, For the Mode in time The description texts are respectively the description text corresponding to the text data, the behavior description text corresponding to the behavior data and the visual description text corresponding to the image data.

[0042] Entity Twin Vector It is a multimodal feature vector generated by fusing semantic vectors, behavioral vectors, visual vectors, text data, and their corresponding descriptive text through a dynamic gating mechanism. It represents the multidimensional dynamic features of an entity in the temporal dimension (including textual semantics, behavioral patterns, visual attributes, etc.). It generates dynamic gating weights through a pre-trained large language model and adaptively adjusts the fusion ratio of each modal information to achieve temporal evolution modeling of entity features. Its significance lies in providing real-time and comprehensive entity representation for knowledge graphs, supporting dynamic focusing of entity features based on business rules, improving the accuracy and flexibility of knowledge graphs in entity association analysis and potential relationship mining, and adapting to entity modeling needs in multiple scenarios such as financial risk control and customer profiling.

[0043] Step S2: construct a knowledge graph using the structured data in the customer profile and the entity twin vector.

[0044] The steps of constructing a knowledge graph using the structured data in the customer profile and the entity twin vectors are as follows: Step S201 : extracting business rules based on the business requirements of the target domain, and obtaining entities according to the structured data of the customer profile and the business rules.

[0045] Extracting business rules based on the business needs of the target domain involves analyzing the business objectives of the target domain (such as financial risk control) and, based on the characteristics of customer profile data, extracting a set of rules for entity generation, feature fusion, and relationship mining. For example, in a financial risk control scenario, this might include a screening rule such as "high-value customers must have a spending amount greater than 100,000 yuan and a credit rating of A or higher," or a feature weighting rule such as "age weighting 0.6, consumption frequency weighting 0.4" for customer segmentation. Business rules can be expressed in natural language or structured formats, such as industry-standard documents.

[0046] Wherein, obtaining an entity according to the structured data of the customer profile and the business rules includes: Step S20101: vectorize the structured data of the customer file to obtain a structured data vector.

[0047] The structured data of customer files includes numerical structured data and categorical structured data.

[0048] The process of vectorizing the structured data of the customer profile to obtain a structured data vector specifically includes the following steps: Step S2010101: normalize the numerical structure data to generate a numerical feature vector.

[0049] In this embodiment, the numerical fields of structured data in the customer file (such as age, consumption amount, transaction frequency, etc.) are first identified. According to the numerical range and business meaning of different fields, the corresponding normalization method is selected (such as the minimum-maximum scaling method to scale the numerical value to the [0,1] interval, or the Z-score normalization method to eliminate the dimension effect). Then, the original data of each numerical field is normalized. Finally, all normalized numerical fields are arranged in sequence to form a numerical feature vector. This vector retains the relative size relationship and distribution characteristics of the numerical data, which is convenient for subsequent weighted fusion and entity feature analysis after splicing with the embedding vector of the categorical structured data.

[0050] Step S2010101: pre-train embedding coding is performed on the category structure data to generate a category feature vector.

[0051] In this embodiment, the categorical fields (such as industry, customer level, and region labels) of the structured data in the customer profile are first identified, and each category value is converted into a corresponding text label (such as "Industry = Finance" and "Region = Shanghai"). The category text labels are then embedded and encoded using a pre-trained generative large language model (such as LLaMA-2 and GPT-4) or a specialized word embedding model (such as Word2Vec and BERT). The discrete category labels are mapped into continuous high-dimensional vectors through the embedding layer of the model. Finally, the embedding vectors of all categorical fields are concatenated in field order to form a category feature vector, which captures the semantic associations and potential structures of the category labels.

[0052] Step S2010101: concatenate the numerical feature vector and the category feature vector to obtain a structural data vector.

[0053] In this embodiment, after completing the normalization processing of the numerical structured data to generate a numerical feature vector (such as a vector containing normalized numerical values ​​such as age and consumption amount), and the pre-trained embedding coding of the categorical structured data to generate a categorical feature vector (such as a vector containing semantic embeddings such as industry and customer level), the elements of the two vectors are connected end to end in a predefined field order (such as numerical fields first and categorical fields later) to form a complete high-dimensional vector. If the dimension of the numerical feature vector is 10 dimensions and the dimension of the categorical feature vector is 50 dimensions, the dimension of the spliced ​​structured data vector is 60 dimensions. This vector integrates the information representation of the numerical and categorical features in the structured data.

[0054] Step S20102: Perform weighted fusion on the structural data vector and the entity twin vector to construct an entity feature matrix.

[0055] Among them, weighted fusion of the structural data vector and the entity twin vector and construction of the entity feature matrix specifically includes the following sub-steps: using a generative large model to perform semantic extraction on the business rules to generate a fusion weight vector; using the fusion weight vector to perform element-by-element weighted summation of the structural data vector and the entity twin vector to obtain a fusion feature vector; and using the fusion feature vector to construct an entity feature matrix.

[0056] In this embodiment, the business rules for the target domain (e.g., natural language descriptions of "age dimension weight 0.6, consumption amount ≥ 100,000 yuan" or structured rule text) are input into a pre-trained generative large model such as LLaMA-2. The model's semantic parsing capabilities identify the feature fields involved in the rules (e.g., "age," "consumption amount") and their corresponding weight thresholds or priorities (e.g., "0.6," "greater than or equal to 100,000 yuan"). The generative large model generates a fused weight vector that matches the feature vector dimensions based on the dimensional mapping relationship between the feature fields in the structured data vector and the entity twin vector (e.g., "age" corresponds to dimensions 1-128 of the structured data vector). For example, if both the structured data vector and the entity twin vector are 768-dimensional, the model outputs a 768-dimensional weight vector, in which dimensions related to rules such as "age" and "consumption amount" are assigned corresponding weight values ​​(e.g., 0.6). Weights for other dimensions are automatically assigned based on the default logic of the business rules or model learning, ultimately resulting in a fused weight vector.

[0057] Ensure that the fusion weight vector is consistent with the structural data vector and the entity twin vector dimensions, where each element of the fusion weight vector corresponds to the weight value of a feature dimension (range 0-1). For each feature dimension , take the first The dimension value is multiplied by the fusion weight vector Dimension weight , and then take the entity twin vector Dimension value multiplied by , add the two product results to get the first fusion feature vector Dimension value, traverse all dimensions according to this rule, and finally form a complete fusion feature vector, which dynamically integrates the static features of structured data and the multimodal dynamic features of entity twin vectors through weight distribution driven by business rules.

[0058] The fused feature vectors are treated as independent row vectors or column vectors and arranged and stacked in the order of the entity's unique identifier (such as customer ID) to form a two-dimensional matrix. If there are N entities and the dimension of each fused feature vector is D, then the dimension of the entity feature matrix is ​​N×D, where each row corresponds to the comprehensive feature representation of an entity and each column corresponds to the global distribution of a specific feature dimension. The matrix integrates the structured data features and twin features of all entities.

[0059] The essence of this embodiment is to parse the semantics of business rules through a generative large model to determine the weights, and perform element-by-element weighted summation of structured data vectors and entity twin vectors, thereby constructing an entity feature matrix that integrates structured attributes and multimodal dynamic features, and realizing comprehensive digital modeling of entities.

[0060] Step S20103: Analyze the entity feature matrix using a generative large model to generate a candidate entity set.

[0061] In this embodiment, the entity feature matrix is ​​input into the pre-trained generative large model LLaMA-2, and the distribution pattern, similarity and association rules of the feature vectors of each entity in the matrix are analyzed through the attention mechanism of the generative large model. The generative large model first converts the numerical features in the matrix into semantic representations (such as mapping high-dimensional vectors into text descriptions such as "high consumption frequency + young customer group"), and then identifies dense areas or anomalies in the feature space based on the entity generation logic implied by the business rules (such as customer segmentation and risk rating), and generates potential candidate entity initial names or initial category labels.

[0062] Leveraging the semantic understanding and feature association capabilities of the generative big model, the entity feature matrix, which combines structured data and twin features, is analyzed. Through clustering, association rule mining, or semantic matching, entity combinations or groups of related entities with potential business value are identified. The candidate entity set may contain different name representations of the same entity (e.g., non-standardized names such as aliases, abbreviations, and historical names). This is because the generative big model has cross-language and cross-representation semantic alignment capabilities, and can recognize that different forms of reference, such as "Huayi," "Huayi Technology Co., Ltd.", and "HY," refer to the same entity. Furthermore, if the entity feature matrix contains multimodal text data containing aliases, the model will aggregate the different names into candidate entities based on semantic associations. However, these candidate entities themselves may not yet have been standardized and require subsequent standardization through the entity disambiguation module.

[0063] Step S20104: introducing a rule engine, utilizing the rule engine to load the business rules and perform screening and standardization on the candidate entity set to obtain a plurality of entities.

[0064] In this embodiment, a rule engine (such as Drools or EasyRule) is first initialized. Business rules for the target domain (e.g., structured or natural language rules in financial risk control, such as "consumption amount greater than 100,000 yuan and credit rating greater than or equal to A" or "number of abnormal behaviors less than or equal to 2 in the past 30 days") are loaded into the engine in the form of JSON, XML, or rule scripts. The engine then traverses the set of candidate entities output by the generative large model, matching each candidate entity's characteristic description (e.g., "top 10% of customers with frequent consumption in East China") with the loaded business rules. Entities that meet the rules are selected through conditional filtering (e.g., numerical comparison, category matching, and logical combination). Finally, standardized name annotations are applied to these selected entities based on the naming conventions pre-set in the rules (e.g., "High-Value Customer - Platinum Level" and "Potential Risk Customer"), resulting in a set of entities that conform to the business logic (e.g., a list of entities with clear labels such as "High Net Worth Customer" and "Quality Supplier"), ensuring the accuracy of entity definitions and business interpretability.

[0065] Standardization processing refers to loading business rules through a rule engine and performing full-process operations such as rule matching, condition filtering, and name calibration on a set of candidate entities to eliminate the ambiguity, redundancy, and non-standardization of entity features and make them comply with the business rules and data governance standards of the target field.

[0066] Step S202: extract the direct relationship between the entities using the business data in the structured data, and associate the direct relationship with the entity twin vector to obtain an explicit relationship.

[0067] In this embodiment, key business fields are extracted from multi-source data through a data parsing module, such as extracting customer ID, counterparty ID, transaction time and amount from bank statements, and extracting parent company ID, subsidiary ID and shareholding ratio from corporate business information. Predefined field mapping rules (such as data dictionaries or mapping tables) are used to accurately map fields such as "customer ID" and "partner ID" to unique entity IDs in the entity feature matrix, and invalid ID records are eliminated through validity verification. A relationship type dictionary is constructed based on industry business rules. For example, in the financial field, relationship types such as "counterparty" (applicable to fund transaction records), "guarantor" (applicable to credit contracts), and "parent-subsidiary" (applicable to corporate relationships with a shareholding ratio greater than 50%) are defined, and data features are associated through conditional expressions (such as "contract type = procurement" and the relationship type is "supplier"). For each business data record, a raw triple is generated based on "entity 1 - relationship type - entity 2." For example, a transaction record (customer ID = 1001, counterparty ID = 1002, transaction type = transfer) corresponds to the triple (1001, counterparty, 1002). Duplicate records are removed by timestamp or transaction serial number, and duplicate relationship records are aggregated by the latest time or cumulative value (such as cumulative transaction amount). When associating entity twin vectors, the corresponding twin vector is retrieved from the entity feature matrix in real time based on the entity ID through a distributed cache or batch query interface. The vectors of entity 1 and entity 2 are then embedded in the relationship representation in JSON format to form an explicit relationship.

[0068] Step S203: Analyze the entity twin vectors corresponding to the entities through a generative big model to obtain potential relationships.

[0069] The following steps are used to analyze the entity twin vectors corresponding to the entities using the generative big model to obtain the potential relationships: Step S20301: Use the generative big model to analyze the entity twin vector corresponding to the entity to obtain the potential hypothesis relationship and confidence.

[0070] In this embodiment, the input entity twin vector is preprocessed in multiple stages: the feature dimension differences are eliminated through normalization technology, the PCA or t-SNE algorithm is used to reduce the high-dimensional vector to the model adaptation dimension, and the time series vector is framed through a sliding window mechanism to capture the evolution pattern of features over time. Subsequently, the model reasoning ability is activated through dynamic prompt engineering: according to the vector modal composition (such as when the text accounts for more than 40%), the prompt word template is automatically selected (such as based on the text semantic "keyword extraction results" and the behavioral feature "time series pattern summary" to infer the potential association between entities), and a small number of sample examples are inserted to guide the model to generate relationship hypotheses that conform to cross-domain logic.

[0071] During the model inference phase, the multi-head attention mechanism is broken down into two layers: intra-modal and inter-modal processing. The first layer focuses on single-modal features (such as noun phrase weights in text vectors or color distribution features in image vectors), extracting local key features. The second layer establishes cross-modal associations (such as the visual semantic alignment of the text "R&D investment" and the image "laboratory scene"), quantifying the strength of feature interactions using an attention weight matrix. The model generates hypotheses based on a hybrid relational knowledge base (integrating common common sense relationships with scalable industry rules). For example, when the "market share" values ​​in the vectors are close and the "product line overlap" is greater than 0.6, the "competitive relationship" inference module is triggered. Combined with the historical relationship distribution of similar feature combinations in the training data (e.g., 82% of cases correspond to competitive relationships), it generates an initial relationship hypothesis and a confidence level.

[0072] In an optional implementation, the confidence level can also be calibrated. Specifically, a Bayesian network model can be introduced, multimodal feature evidence (such as text evidence, behavioral evidence, and image evidence) can be used as network nodes, and the joint probability can be calculated through a conditional probability table to perform a posteriori correction on the confidence level.

[0073] Step S20302: construct a predefined rule base, and identify the potential hypothesized relationship based on the predefined rule base using the confidence level to obtain a potential relationship.

[0074] In this embodiment, for target areas (such as financial risk control, customer profiling, etc.), business logic rules, industry standard rules and data validation rules are integrated to form a structured rule set, where business logic rules include confidence thresholds for relationship establishment (such as "cross-industry association relationships must have a confidence level greater than 0.8"), feature matching conditions (such as "related entities must have high-frequency transaction records"); industry standard rules cover domain-specific constraints (such as "parent-subsidiary relationships must have a shareholding ratio greater than or equal to 50%", "partners must have industry access qualifications"), and data validation rules involve multimodal feature consistency requirements (such as "potential relationships must match at least two modal features", "time series data must conform to business cycle laws"). All rules are stored in a rule base in a structured form such as JSON, XML or rule scripts to form a predefined rule base, The potential hypothetical relationships and confidence levels output by the generative large model (e.g., "Customer A and Customer B have a supply chain relationship, confidence level 0.72") are entered into the rule base, and the conditions in the rule base are matched one by one. For each potential relationship, its confidence level is checked to see if it meets the rule's preset threshold (e.g., 0.7), and the relationship description is verified to conform to the business logic (e.g., "supply chain relationship" requires the presence of high-frequency transaction records in the corresponding entity's behavior vector). Finally, only relationships that meet both the confidence level requirements and the rule conditions are retained as potential relationships (e.g., "supply chain relationship" with a confidence level of 0.75 and the presence of transaction records). Hypotheses that do not conform to the rules are eliminated (e.g., "partnership relationship" with a confidence level of 0.6), ensuring that the potential relationships possess both data-driven credibility and conform to the logical constraints of the business domain.

[0075] Step S204: construct a knowledge graph based on the entities, the explicit relationships, and the potential relationships.

[0076] In this embodiment, entities filtered and annotated by business rules are used as nodes in the knowledge graph, and each node is attached with its corresponding entity twin vector (including multimodal dynamic features such as text semantics, behavioral patterns, and visual attributes). Explicit relationships of entity twin vectors are extracted and associated from structured business data, as well as potential relationships identified through generative large model analysis and rule base screening as edges of the graph. Each edge is annotated with the relationship type (such as "transaction" or "supply chain"), confidence level (only for potential relationships), and the associated entity dynamic features. Nodes and edges are structurally connected according to the association logic between entities to form a knowledge graph containing entity nodes, explicit edges, potential edges, and multidimensional attributes. This graph supports real-time evolution as entity twin vectors are updated, dynamically reflecting the temporal changes of entities and their associations, and providing a structured knowledge representation and multidimensional reasoning foundation for scenarios such as customer profile analysis and business association mining.

[0077] The knowledge graph is centered on entity twin vectors and time-sensitive edge relationships, covering the full-process dynamics of entity feature evolution, relationship weight update, and graph structure adjustment.

[0078] Step S3: The customer profile includes an interaction log, and the interaction log is used to calculate the edge relationship weight of the knowledge graph.

[0079] Calculating the edge relationship weights of the knowledge graph using the interaction log specifically includes the following sub-steps: Step S301: Add the time information in the interaction log to the knowledge graph to form a time series knowledge graph.

[0080] In this implementation, we first extract the timestamp field and the corresponding entity interaction event (such as "Customer A and Customer B had a transaction on 2025-06-01") from the interaction log (such as customer behavior logs, transaction record logs, etc.), and associate the event with the entity node in the knowledge graph through the entity ID mapping rule. Then, we add time attributes to the explicit and potential relationships in the graph. For example, we record fields such as "relationship establishment time" and "last update time" in the relationship edges. For periodic or continuous relationships (such as long-term cooperation agreements), we mark the time interval (start time - end time). Finally, we introduce the time dimension into the knowledge graph structure, and construct a timestamp entity state transition chain (such as the node sequence of entity A's credit rating changing over time). This forms a temporal knowledge graph with time as the main axis and entities and relationships as nodes and edges. This supports functions such as time slice query and temporal association analysis, and realizes full-cycle tracking of the evolution process of entities and their relationships.

[0081] The time series knowledge graph is an extension of the time dimension of the knowledge graph. It constructs a timestamp graph structure through the time information of the interaction log and supports relationship analysis by time slices.

[0082] Step S302: extracting time series data of edge relationships in the time series knowledge graph to form a time series vector.

[0083] In this embodiment, all edges in the temporal knowledge graph (including explicit and latent relationships) are traversed. For each edge's time attribute (such as the timestamp of a single interaction record or the time interval of a persistent relationship), time series data is extracted by entity pairs (such as "entity A-entity B"). For instantaneous edge relationships (such as a single transaction), the numerical representation of their timestamps (such as a Unix timestamp or a numerical encoding converted to "year-month-day") is extracted and arranged in chronological order to form a one-dimensional time series. For persistent edge relationships (such as long-term cooperation agreements), features such as the duration of the relationship (such as the number of days from the end time to the start time) and the starting point of the time interval are calculated and converted into a numerical vector. Finally, all time series data is standardized (such as normalized to the [0,1] interval or Z-score normalization) and spliced ​​by entity pair dimension to form a time series vector. This vector captures the temporal distribution pattern and evolution characteristics of the relationship between entities.

[0084] Step S303: construct a weight calculation model based on a time series graph neural network, and obtain edge relationship weights using the weight calculation model according to the time series vector and the semantic vector.

[0085] In this embodiment, historical customer files before a predetermined time (such as two years) are used, and a large model is used to extract semantic vectors from historical text data, behavioral vectors from behavioral data, and visual vectors from image data. The behavioral and visual vectors are converted into historical description texts, and are fused through a dynamic gating mechanism to generate historical entity twin vectors. Secondly, timestamps are extracted from interaction logs two years ago, and historical time series data are grouped by entity pairs. After standardization and sine-cosine position encoding, historical time series vectors are generated. The historical time vector is extracted from the historical time series vector to obtain a historical time coding vector. At the same time, the semantic vector is extracted from the historical entity twin vector, and the two are spliced ​​into a composite feature vector. Finally, based on the business rules and expert experience in the historical structured data, the explicit relationship weights are labeled (such as "historical transaction amount ≥ 1 million" is labeled as 1.0). The dynamic characteristics of the historical entity twin vectors are analyzed by the generative large model to obtain the potential hypothetical relationship and confidence. After screening with the historical predefined rule library, it is used as the potential relationship label (0-1 value). Finally, with "historical time series features + historical semantic features" as input features and historical edge weight labels as supervision signals, the training samples are constructed, and then the weight calculation model is trained using the training samples to obtain a pre-trained weight calculation model. Finally, the time series vector and semantic vector are input into the weight calculation model to obtain the edge relationship weight.

[0086] In an optional embodiment, the weight calculation model can also be improved. By setting a time decay module in the input layer of the weight calculation model, the historical interaction data can be weighted by temporal importance, the influence of long-term interactions can be suppressed, and the weight of recent interactions can be strengthened, so that the graph can dynamically reflect business changes. By setting a self-attention layer after the input layer, the association weights between the time series vector and the semantic vector can be explicitly calculated, solving the black box problem of the original model relying on the neural network to implicitly learn feature associations, making the edge relationship weights more explainable and in line with business logic, thereby improving the knowledge graph's ability to mine implicit relationships.

[0087] Constructing a weight calculation model based on a time-series graph neural network includes: setting a time decay module in the input layer of the weight calculation model; setting a self-attention layer after the input layer; and constructing a weight calculation model based on the time decay module and the self-attention layer.

[0088] The time decay module satisfies the following formula: in, For Entity and entities The decayed time series vector, is the total number of interactions between entities, For the The time decay coefficient of the interaction, For the The original time series vector of interactions, is the natural exponential function, is the empirical attenuation coefficient, For the The time interval between the interaction and the current time.

[0089] The self-attention layer satisfies the following formula: in, For Entity and entities The fusion feature vector of is the self-attention weight, For Entity and entities The vector after linear projection of the decayed time series vector, For Entity and entities The vector after linear projection of the semantic vector, is the dimension of the projected feature vector.

[0090] The above empirical attenuation coefficient The value of is determined by grid search, and the empirical attenuation coefficient is preset The candidate value set is used to calculate the Spearman correlation coefficient between edge weight and actual business intensity on historical data, and the one with the highest correlation coefficient is selected. Value, such as in financial risk control scenarios When the weight of recent transactions is attenuated, it is more in line with the law of risk transmission.

[0091] Step S4: Adjust the knowledge graph based on the edge relationship weights.

[0092] The step of adjusting the knowledge graph based on the edge relationship weights specifically includes the following sub-steps: Step S401: setting a weight threshold, and comparing the weight threshold with the edge relationship weight; In this embodiment, a weight threshold is set according to business requirements or domain rules (such as the critical value of association strength in risk assessment and the credibility threshold of the recommendation system), and the relationship weight value of each edge output by the time series graph neural network model is compared with the threshold one by one.

[0093] Step S402: Update or delete the edge in the knowledge graph corresponding to the edge relationship weight according to the comparison result.

[0094] In this embodiment, edges with weight values ​​greater than or equal to a threshold are determined to have strong associations and are retained (e.g., marked as "high confidence association"). Edges with weight values ​​less than the threshold are determined to have weak associations and are deleted.

[0095] like Figure 2 As shown, on the other hand, the present invention also provides a customer profile knowledge graph construction system based on large model driving, including: a processor, an input device, an output device and a memory, wherein the processor, the input device, the output device and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute relevant steps of a relevant embodiment of a customer profile knowledge graph construction method based on large model driving of the present invention.

[0096] The present invention provides a large-model-driven customer profile knowledge graph construction system. Each functional component can be integrated into a single processing component, each component can exist physically separately, or two or more components can be integrated into a single component. These integrated components can be implemented as either hardware or software functions.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A method for constructing a customer profile knowledge graph based on a large model, characterized in that: The method comprises: Encode customer files using large model technology, and fuse the encoding results to generate entity twin vectors; Constructing a knowledge graph using the structured data in the customer profile and the entity twin vector; The customer profile includes an interaction log, and the edge relationship weights of the knowledge graph are calculated using the interaction log; The knowledge graph is adjusted based on the edge relationship weights.

2. According to claim 1, a method for constructing a customer profile knowledge graph based on a large model drive is characterized in that: The customer profile includes text data, behavior data, and image data. The method of encoding the customer profile using the large model technology and fusing the encoding results to generate entity twin vectors includes: Using large model technology to extract semantic features from the text data, extract behavioral features from the behavioral data, and extract visual features from the image data, to obtain semantic vectors, behavioral vectors, and visual vectors; Converting the semantic vector, the behavioral vector, and the visual vector into a semantic description text, a behavioral description text, and a visual description text, respectively; A dynamic gating mechanism is used to fuse the semantic vector, the behavioral vector, the visual vector, the semantic description text, the behavioral description text and the visual description text to obtain an entity twin vector.

3. According to claim 2, a method for constructing a customer profile knowledge graph based on a large model drive is characterized in that: The entity twin vector satisfies the following formula: in, Over time Updated entity twin vector, is the Sigmoid function, For pre-trained generative large language models, For the Mode in time Description text of For the Mode in time The eigenvector of .

4. The method for constructing a customer profile knowledge graph based on a large model drive according to claim 1, characterized in that: The step of constructing a knowledge graph using the structured data in the customer profile and the entity twin vector includes: Extracting business rules based on business requirements of the target domain, and obtaining entities based on the structured data of the customer profile and the business rules; Extracting direct associations between the entities using business data in the structured data, and associating the direct associations with the entity twin vectors to obtain explicit relationships; Analyze the entity twin vectors corresponding to the entities through a generative big model to obtain potential relationships; A knowledge graph is constructed based on the entities, the explicit relationships, and the latent relationships.

5. According to claim 4, a method for constructing a customer profile knowledge graph based on a large model drive is characterized in that: The obtaining of an entity according to the structured data of the customer profile and the business rules includes: Vectorizing the structured data of the customer profile to obtain a structured data vector; Performing weighted fusion on the structure data vector and the entity twin vector to construct an entity feature matrix; Analyzing the entity feature matrix using a generative large model to generate a set of candidate entities; A rule engine is introduced, and the business rules are loaded by using the rule engine and the candidate entity set is screened and standardized to obtain a plurality of entities.

6. The method for constructing a customer profile knowledge graph based on a large model drive according to claim 5, characterized in that: The structured data of the customer file includes numerical structured data and categorical structured data. The vectorization of the structured data of the customer file to obtain the structured data vector includes: Normalizing the numerical structure data to generate a numerical feature vector; Performing pre-training embedding coding on the categorical structure data to generate a categorical feature vector; The numerical feature vector and the category feature vector are concatenated to obtain a structural data vector.

7. The method for constructing a customer profile knowledge graph based on a large model drive according to claim 4, characterized in that: The entity twin vectors corresponding to the entities are analyzed by the generative big model to obtain potential relationships including: Analyze the entity twin vector corresponding to the entity using a generative big model to obtain a potential hypothesis relationship and confidence level; A predefined rule base is constructed, and the confidence level is used to identify the potential hypothesized relationship based on the predefined rule base to obtain the potential relationship.

8. According to claim 2, a method for constructing a customer profile knowledge graph based on a large model drive, characterized in that: The step of calculating the edge relationship weight of the knowledge graph using the interaction log includes: Adding the time information in the interaction log to the knowledge graph to form a time series knowledge graph; Extracting time series data of edge relationships in the time series knowledge graph to form a time series vector; A weight calculation model is constructed based on a time series graph neural network, and the edge relationship weight is obtained using the weight calculation model according to the time series vector and the semantic vector.

9. The method for constructing a customer profile knowledge graph based on a large model drive according to claim 1, characterized in that: The adjusting the knowledge graph based on the edge relationship weights includes: Setting a weight threshold, and comparing the weight threshold with the edge relationship weight; According to the comparison result, the edge in the knowledge graph corresponding to the edge relationship weight is updated or deleted.

10. A customer profile knowledge graph construction system based on a large model drive, characterized by: include: A processor, an input device, an output device and a memory, wherein the processor, the input device, the output device and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute any one of claims 1 to 9 of the method for constructing a customer profile knowledge graph based on a large model drive.

Images