A large model driving-based customer profile knowledge graph construction method and system

Abstract

The application relates to the field of knowledge graph construction, in particular to a customer archive knowledge graph construction method and system based on a large model drive. The method comprises the following steps: coding a customer archive by using a large model technology, and fusing the coding results to generate entity twin vectors; constructing a knowledge graph by using structured data in the customer archive and the entity twin vectors; the customer archive comprises an interaction log, the edge relationship weight of the knowledge graph is calculated by using the interaction log; and the knowledge graph is adjusted based on the edge relationship weight. The application solves the problem that a traditional knowledge graph in the prior art cannot accurately capture entity dynamic characteristics and hidden correlations, so that the constructed knowledge graph cannot meet dynamic business requirements such as real-time risk control and accurate marketing. The application focuses on the accuracy and simplicity of knowledge graph construction while deeply mining entity relationships, and greatly improves the explainability and reasoning ability of the knowledge graph.

Description

TECHNICAL FIELD

[0001] The application relates to the field of knowledge graph construction, and in particular to a customer archive knowledge graph construction method and system based on a large model drive. BACKGROUND

[0002] As a structured knowledge representation method, a knowledge graph has wide application prospects in customer archive management and business correlation analysis. Traditional customer archive management is mostly based on a relational database, and can only store entity static attributes and explicit correlations, and cannot capture entity dynamic characteristics and implicit correlations, resulting in insufficient dynamicity and reasoning capability of the knowledge graph.

[0003] With the development of big data technology, customer archive data has the characteristics of multi-modal, time sequence and high dimension. Traditional graph construction methods based on a rule engine or a shallow neural network face many challenges, such as inefficient multi-modal feature fusion, text semantic ambiguity, behavior time sequence dependency and difficulty in effectively fusing image visual features through traditional feature engineering, resulting in single entity representation dimension, inability to comprehensively depict customer dynamic characteristics, limited implicit relationship mining capability, dependence on pre-defined rules or statistical correlations, difficulty in capturing cross-modal deep semantic correlations and generating potential relationship hypotheses, inability to update entity characteristics and relationship weights in real time, and difficulty in supporting real-time risk control and precise marketing business requirements.

[0004] Therefore, there is an urgent need for a knowledge graph construction method that can deeply mine the relationships between entities, so that the knowledge graph can better depict multi-dimensional entities. SUMMARY

[0005] In view of the defects in the prior art, the application provides a customer archive knowledge graph construction method and system based on a large model drive, which solves the problem that the traditional knowledge graph in the prior art cannot accurately capture entity dynamic characteristics and implicit correlations, so that the constructed knowledge graph cannot meet the business requirements of real-time risk control and precise marketing.

[0006] To achieve the above purpose, the application provides a customer archive knowledge graph construction method based on a large model drive, which comprises the following steps: encoding a customer archive by using a large model technology, and fusing the encoding results to generate an entity twin vector; constructing a knowledge graph by using structured data in the customer archive and the entity twin vector; the customer archive comprises an interaction log, the edge relationship weight of the knowledge graph is calculated by using the interaction log; and the knowledge graph is adjusted based on the edge relationship weight.

[0007] This invention utilizes large-scale modeling technology to perform multimodal encoding and fusion of customer profiles, generating entity twin vectors that evolve over time. This overcomes the limitations of traditional methods that rely on static and singular entity features, dynamically capturing real-time changes in customer semantics, behavioral patterns, and visual attributes. By constructing a knowledge graph using customer profiles and twin vectors, it achieves deep binding between explicit relationships (such as transaction records and contractual relationships) and entity features, while providing a multi-dimensional feature foundation for latent relationship mining. By calculating edge relationship weights using interaction logs, the time dimension is incorporated into the graph modeling, allowing the strength of relationships between entities to quantify temporal patterns such as interaction frequency and time intervals, thus improving the graph's ability to depict the dynamic evolution of business processes.

[0008] Optionally, the customer profile includes text data, behavioral data, and image data. Encoding the customer profile using large model technology and fusing the encoded results to generate entity twin vectors includes: extracting semantic features from the text data, extracting behavioral features from the behavioral data, and extracting visual features from the image data using large model technology to obtain semantic vectors, behavioral vectors, and visual vectors; converting the semantic vectors, behavioral vectors, and visual vectors into semantic description text, behavioral description text, and visual description text, respectively; and fusing the semantic vectors, behavioral vectors, visual vectors, semantic description text, behavioral description text, and visual description text using a dynamic gating mechanism to obtain entity twin vectors.

[0009] This invention utilizes a large model to perform specialized feature extraction for different modal data characteristics, achieving deep decoupling and accurate characterization of multimodal features, thus avoiding the limitations of traditional feature engineering. Behavioral and visual vectors are transformed into natural language descriptive text, and the semantic generation capabilities of the large model eliminate the multimodal semantic gap, laying a unified semantic foundation for cross-modal fusion. A dynamic gating mechanism is employed to fuse multimodal data, 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:

[0011]

[0012] in, For time Updated entity twin vector, For the Sigmoid function, For pre-trained generative large language models, For the first Each mode in time Description text, For the first A modal at a time characteristic vector.

[0013] The application processes multi-modal description text by generating large language model, dynamically weights and fuses each modal characteristic vector by combining Sigmoid function, realizes entity representation update evolving with time, not only uses semantic understanding ability of generating large language model to mine multi-modal data value, but also accurately adapts importance of each modal under different time by dynamic gating mechanism, and further improves accuracy of entity twin vectors.

[0014] Optionally, the constructing a knowledge graph by using the structured data in the customer profile and the entity twin vectors comprises: extracting business rules based on business requirements of a target field, and obtaining entities according to the structured data of the customer profile and the business rules; extracting direct association relationships between the entities by using business data in the structured data, and associating the direct association relationships with the entity twin vectors to obtain explicit relationships; analyzing the entity twin vectors corresponding to the entities by a generating large model to obtain potential relationships; and constructing a knowledge graph based on the entities, the explicit relationships and the potential relationships.

[0015] The application takes target field business requirements as an anchor, accurately extracts entities by combining customer profile structured data and business rules, and ensures that the entities are consistent with business logic. Direct associations between entities are mined by using business data, and explicit relationships are formed by associating entity twin vectors, so that entity associations not only retain direct contacts in business scenarios, but also integrate dynamic characteristics of entities, making relationship expression more rich and stereoscopic. Potential relationships are mined by analyzing dynamic characteristics of entity twin vectors with the help of a generating large model, breaking through the limitations of traditional explicit business data, and discovering relationships that are not easily perceived, such as customer behavior trend association and cross-scenario implicit interaction, expanding the depth and breadth of the knowledge graph. The knowledge graph constructed by the cooperation of entities, explicit relationships and potential relationships not only consolidates business basic associations, but also extends and mines implicit values, improving the knowledge driving ability of the knowledge graph.

[0016] Optionally, the obtaining entities according to the structured data of the customer profile and the business rules comprises: vectorizing the structured data of the customer profile to obtain a structured data vector; weighting and fusing the structured data vector and the entity twin vectors to construct an entity feature matrix; analyzing the entity feature matrix by a generating large model to generate a candidate entity set; introducing a rule engine, loading the business rules by the rule engine, and screening and standardizing the candidate entity set to obtain a plurality of entities.

[0017] Optionally, the structured data of the customer profile includes numerical structured data and category structured data, and the vectorization of the structured data of the customer profile to obtain a structured data vector includes: normalizing the numerical structured data to generate a numerical feature vector; pre-training embedding coding is performed on the category structured data to generate a category feature vector; and the numerical feature vector and the category feature vector are spliced to obtain the structured data vector.

[0018] The structured data is vectorized and fused with entity twin vectors to construct a feature matrix, the advantages of multi-source data are fully integrated, and the entity features are more comprehensive. With the help of a generative large model, candidate entities are generated, potential entity information is mined, and the limitations of traditional methods are broken through. Then, a rule engine is introduced to filter and standardize candidate entities according to business rules, which not only ensures that the entities meet business requirements, but also standardizes the form of the entities, thereby improving the quality and consistency of the entities.

[0019] Optionally, the analysis of the entity twin vector corresponding to the entity by the generative large model to obtain a potential relationship includes: analyzing the entity twin vector corresponding to the entity by the generative large model to obtain a potential hypothesis relationship and a confidence level; and constructing a predefined rule library, and based on the predefined rule library, the potential hypothesis relationship is discriminated by using the confidence level to obtain a potential relationship.

[0020] The deep analysis of the entity twin vector by the generative large model can capture cross-modal and time-series semantic associations that are difficult to discover by traditional methods, generate potential hypothesis relationships and quantify confidence levels, and inject new knowledge discovery capabilities into the graph. The introduction of the predefined rule library effectively filters low-quality or non-domain-knowledge-compliant hypothesis relationships through business logic constraints, which is consistent with business reality and improves the scientificity and accuracy of potential relationship mining.

[0021] Optionally, the calculation of the edge relationship weight of the knowledge graph by using the interaction log includes: adding time information in the interaction log into the knowledge graph to form a time-series knowledge graph; extracting 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 by using the weight calculation model according to the time-series vector and the semantic vector.

[0022] The application upgrades the knowledge graph into a time sequence knowledge graph by introducing time information of the interaction log, and realizes time sequence dynamic modeling of the edge relationship weight.

[0023] Optionally, the adjusting the knowledge graph based on the edge relationship weight comprises: setting a weight threshold, comparing the weight threshold with the edge relationship weight, and updating or deleting an edge corresponding to the edge relationship weight in the knowledge graph according to a result of the comparison.

[0024] By setting the weight threshold and comparing the edge relationship weight, the application can dynamically filter invalid or weakly associated edges and retain strongly associated relationships. This mechanism ensures that the knowledge graph always focuses on key business associations and avoids interference from redundant information, thereby improving the simplicity and readability of the knowledge graph.

[0025] Another aspect of the application also provides a customer profile knowledge graph construction system based on a large model driver, 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 connected to each other, and the memory is used to store a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute the customer profile knowledge graph construction method based on the large model driver according to any one of the previous aspect.

[0026] The customer profile knowledge graph construction system based on the large model driver has a compact structure, stable performance, high integration and simple composition, and can stably execute the customer profile knowledge graph construction method based on the large model driver provided in the previous aspect, thereby further improving the overall applicability and practical application ability of the application. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A customer profile knowledge graph construction system based on a large model driver according to an embodiment of the application is shown in the flowchart.

[0028] Figure 2 A customer profile knowledge graph construction system based on a large model driver according to an embodiment of the application is shown in the structural schematic diagram. DETAILED DESCRIPTION

[0029] Specific embodiments of the present application will now be described in detail with reference to the drawings, which are provided by way of example and not limitation. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one of ordinary skill in the art that the present application can be practiced without the specific details. In other instances, well-known circuits, software or methods have not been described in detail in order to avoid obscuring the present application. It will be appreciated that the various units shown in the figures can be combined in a system-on-a-chip (SoC) along with other units or they can be provided separately.

[0030] Reference throughout this specification to "an embodiment", "embodiments", "one example", or "an example" means that a particular feature, structure, or characteristic described in connection with the embodiment or example is included in at least one embodiment of the application. Therefore, appearances of the phrases "in one embodiment", "in embodiments", "one example" or "an example" in various places throughout this specification are not necessarily referring to the same embodiment or example. Furthermore, the particular features, structures, or characteristics can be combined in any suitable

[0031] In an optional embodiment, referring to Figure 1 As shown in a large model driven customer profile knowledge graph construction method, comprising the following steps: Figure 1

[0032] Step S1, using large model technology to encode customer profile, and fusing the results of the encoding to generate entity twin vectors.

[0033] Among them, using large model technology to encode customer profile, and fusing the results of the encoding to generate entity twin vectors specifically includes the following steps:

[0034] Step S101, using large model technology to extract semantic features from the text data, behavior features from the behavior data, and visual features from the image data, respectively, to obtain semantic vectors, behavior vectors, and visual vectors.

[0035] In this embodiment, the customer profile includes text data, behavior data, and image data, wherein the customer profile is the profile data within a predetermined time (such as two years). The structured data of the customer profile usually contains a unique identifier ID, and the pre-classified Chinese text data, behavior data, and image data of the customer profile can be classified by the unique identifier ID, so the semantic vectors, behavior vectors, and visual vectors obtained after feature extraction can be associated by the unique identifier ID.

[0036] ​The semantic feature extraction of text data using large model technology obtains a semantic vector. First, the text data is preprocessed, the text is converted into a token sequence by a tokenizer, and a start symbol and an end symbol are added. The model input length is truncated or padded to adapt to the model input length. Then, a pre-trained generative large language model such as LLaMA-2 or GPT-4 is selected, and the preprocessed token sequence is input into the generative large language model. The context representation of each token is calculated through the Transformer layer of the generative large language model. Then, the vector of a specific token is extracted from the last hidden state 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, different sizes of pre-trained models can be selected according to task requirements, and further optimization of semantic extraction effect in specific fields can be achieved through fine-tuning.

[0037] The behavior feature extraction of behavior data using large model technology obtains a behavior vector. First, the time series feature engineering of behavior data (such as user click, transaction, browsing, etc. sequence data) is performed, which is converted into a standardized time series containing timestamp, behavior type, and numerical indicators, and basic features such as time interval, behavior frequency, and trend change are extracted. Then, a pre-trained time series large model is used, and the time series feature sequence is input into the model. The time dependence, periodic pattern, and abnormal fluctuation features of the behavior data are captured through the attention mechanism or recurrent neural network layer of the model. The feature vectors at each time step output by the model are aggregated into a fixed-dimensional behavior vector 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 behavior data to behavior vector.

[0038] The visual feature extraction of image data using large model technology obtains a visual vector. First, the image data is preprocessed, including scaling to standard size, normalizing pixel values, random cropping or horizontal flipping to enhance data diversity, and targeted processing according to image type (such as contrast adjustment). Then, a pre-trained visual large model (such as CLIP, Swin Transformer, or ResNet-50 convolutional neural network) is selected, and the preprocessed image is input into the model. The local features (such as texture, edge) and global semantic information (such as object category, spatial relationship) of the image are automatically extracted through the convolutional layers of the model. 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.

[0039] The creation time of the text data, the interaction time of the behavior data, and the shooting time of the image data are uniformly converted into Unix timestamps with a precision of seconds. For data with a time deviation exceeding a preset threshold (e.g., 1 hour), a linear interpolation method is used for correction (e.g., for an image with a shooting time 2 hours different from the corresponding transaction behavior time, an intermediate time feature vector is generated by interpolation according to the time sequence), so as to realize cross-modal time alignment.

[0040] In step S102, the semantic vector, the behavior vector, and the visual vector are respectively converted into semantic description text, behavior description text, and visual description text.

[0041] In this embodiment, for the semantic vector, a Seq2Seq model fine-tuned on industry text (such as contracts and interaction logs) is used, combined with a prompt word engineering (such as “extract contract terms, interaction demands, and sentiment orientation”), to decode the business topics, key entities, term demands, and other features implied in the semantic vector into structured natural language descriptions (for example, “In June 2025, customer A signed a three-year framework agreement with supplier B, with a payment period of 45 days”). The text fluency is optimized through beam search, and the time anchor point and subject identifier are unified with the behavior and visual description texts, realizing the alignment of multi-modal features in the time and business dimensions. For the behavior vector, a pre-trained time series model T5-TimeSeries is used to decode the time interval, behavior frequency, trend change, and other behavior feature vectors in the behavior vector, and a natural language description is generated through the Transformer layer of T5-TimeSeries. For example, the behavior vector of “high-frequency clicks + consumption amount fluctuation increase” is converted into the behavior description text of “the user clicks on the product detail page more than 5 times a day in the past 30 days, and the consumption amount increases by 40% compared with last month”. For the visual vector, a pre-trained visual language model such as BLIP, CLIP, or an image description generation model Show and Tell is used to map the color, texture, object category, and other image semantic features captured in the vector into text. For example, the visual vector of “high-rise night view + traffic light” is converted into the visual description text of “urban night scene with high-rise buildings and dense traffic on the road”. The above conversion process can guide the model to generate structured descriptions (such as short sentences containing time, location, and behavior type) through prompt engineering, and optimize the text fluency and information accuracy through decoding strategies such as beam search.

[0042] In step S103, a dynamic gating mechanism is used to fuse the semantic vector, the behavior vector, the visual vector, the semantic description text, the behavior description text, and the visual description text to obtain an entity twin vector.

[0043] 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 using the Sigmoid function to reflect the importance of each modality at the current time. The normalized gating coefficients are then weighted element-wise with the feature vector of the corresponding modality, and the weighted results of all modalities are summed to obtain the entity twin vector that evolves over time.

[0044] Generative large language models learn the temporal importance of features across modalities during training and inference phases. They construct training sets using historical multimodal data (with timestamps) and combine modal descriptions of the same entity at different points in time. As input, each modality is labeled with its business importance at the corresponding time (e.g., in a financial scenario, the semantic vector weight of "quarterly financial report text" should be higher than that of daily behavioral data). Supervised learning is used to optimize the model's ability to represent temporal importance. (Regarding time...) Input Generative large language models output weight vectors with dimensions consistent with modal feature vectors, which are then normalized using the Sigmoid function to obtain gating coefficients. When a certain mode is in time When new data (such as sudden trading activity) triggers a preset incremental threshold (e.g., the data volume exceeds the historical average by 20%), the gating weight will be updated in real time to ensure... Prioritize the integration of key features.

[0045] The entity twin vector satisfies the following formula:

[0046]

[0047] in, For time Updated entity twin vector, For the Sigmoid function, For pre-trained generative large language models, For the first Each mode in time Description text, For the first Each mode in time eigenvectors.

[0048] In this embodiment, For the first Each mode in time The descriptive texts are the descriptive texts corresponding to text data, the behavioral descriptive texts corresponding to behavioral data, and the visual descriptive texts corresponding to image data.

[0049] Entity twin vector The entity twin vector is a multi-modal feature vector generated by fusing semantic vectors, behavior vectors, visual vectors, text data, and corresponding description texts through a dynamic gating mechanism. It represents the multi-dimensional dynamic characteristics of an entity in the time dimension, including text semantics, behavior patterns, visual attributes, etc. A pre-trained large language model is used to generate dynamic gating weights to adaptively adjust the fusion ratio of each modal information, thereby modeling the time evolution of entity features. The significance of the entity twin vector is to provide real-time and comprehensive entity representation for a knowledge graph, support dynamic focusing of entity features based on business rules, improve the accuracy and flexibility of entity correlation analysis and potential relationship mining in the knowledge graph, and adapt to entity modeling needs in multiple scenarios such as financial risk control and customer profiling.

[0050] Step S2, constructing a knowledge graph using the structured data in the customer profile and the entity twin vector.

[0051] The step of constructing a knowledge graph using the structured data in the customer profile and the entity twin vector includes the following steps:

[0052] Step S201, extracting business rules based on the business requirements of the target field, and obtaining entities according to the structured data of the customer profile and the business rules.

[0053] Extracting business rules based on the business requirements of the target field means analyzing the business goals of the target field (such as financial risk control), combining the characteristics of the customer profile data, and refining a rule set for entity generation, feature fusion, and relationship mining. For example, in the financial risk control scenario, a screening rule is extracted that "high-value customers must meet the condition of consumption amount greater than 100,000 yuan and credit rating greater than or equal to A level", or in the customer segmentation scenario, a feature weighting rule is extracted that "age dimension weight 0.6, consumption frequency dimension weight 0.4". Business rules can be expressed in natural language or structured format through industry specification documents.

[0054] The step of obtaining entities according to the structured data of the customer profile and the business rules includes:

[0055] Step S20101, vectorizing the structured data of the customer profile to obtain a structured data vector.

[0056] The structured data of the customer profile includes numerical structured data and categorical structured data.

[0057] The step of vectorizing the structured data of the customer profile to obtain a structured data vector includes the following steps:

[0058] Step S2010101, normalizing the numerical structured data to generate a numerical feature vector.

[0059] In this embodiment, first, the numerical fields of structured data in the customer profile (such as age, consumption amount, transaction frequency, etc.) are identified, and according to the numerical range and business meaning of different fields, the corresponding normalization method (such as the min-max scaling method to scale the values to the [0, 1] interval, or the Z-score standardization method to eliminate the dimension effect) is selected, then the original data of each numerical field is normalized, and finally all the normalized numerical fields are arranged in order to form a numerical feature vector, which retains the relative size relationship and distribution characteristics of the numerical data, facilitating subsequent weighted fusion and entity feature analysis after splicing with the embedding vector of the categorical structured data.

[0060] Step S2010101, pre-training embedding encoding is performed on the categorical structured data to generate a category feature vector.

[0061] In this embodiment, first, the categorical fields of structured data in the customer profile (such as industry, customer level, regional label, etc.) are identified, and each category value is converted into a corresponding text label (such as “industry = finance” and “region = Shanghai”); then a pre-trained generative large language model (such as LLaMA-2, GPT-4) or a special word embedding model (such as Word2Vec, BERT) is used to perform embedding encoding on the category text label, and the embedding layer of the model is used to map the discrete category label to a continuous high-dimensional vector, and finally the embedding vectors of all categorical fields are spliced in order to form a category feature vector, which captures the semantic association and potential structure of the category label.

[0062] Step S2010101, the numerical feature vector and the category feature vector are spliced to obtain a structured data vector.

[0063] In this embodiment, after completing the normalization of numerical structured data to generate a numerical feature vector (such as a vector containing normalized values such as age and consumption amount), and pre-training embedding encoding of categorical structured data to generate a category feature vector (such as a vector containing semantic embeddings such as industry and customer level), the elements of the two vectors are sequentially connected end-to-end according to the pre-defined field order (such as numerical fields first and then categorical fields), to form a complete high-dimensional vector. If the numerical feature vector has a dimension of 10 and the category feature vector has a dimension of 50, the spliced structured data vector has a dimension of 60, which integrates the information representation of numerical and categorical features in structured data.

[0064] Step S20102, the structured data vector and the entity twin vector are weighted and fused to construct an entity feature matrix.

[0065] The weighting and fusing of the structure data vector and the entity twin vector and constructing the entity feature matrix specifically comprises the following sub-steps: performing semantic extraction on the business rules by using a generative large model to generate a fusion weight vector; performing element-by-element weighted summation on the structure data vector and the entity twin vector by using the fusion weight vector to obtain a fusion feature vector; and constructing an entity feature matrix by using the fusion feature vector.

[0066] In the present embodiment, the business rules of the target field (such as the natural language description "age dimension weight 0.6, consumption amount ≥ 10 million yuan" or the structured rule text) are input into a pre-trained generative large model such as LLaMA-2, and the semantic analysis capability of the model is used to identify the feature fields (such as "age", "consumption amount") involved in the rules and their corresponding weights or priorities (such as "0.6", "greater than or equal to 10 million yuan"). The generative large model generates a fusion weight vector matching the dimension of the feature vector according to the dimension mapping relationship of the feature fields in the structure data vector and the entity twin vector (such as "age" corresponding to the 1-128th dimension of the structure data vector). For example, if the structure data vector and the entity twin vector are both 768-dimensional, the model outputs a 768-dimensional weight vector, wherein the dimensions related to the rules such as "age", "consumption amount" are assigned corresponding weight values (such as 0.6), and the weights of other dimensions are automatically assigned according to the default logic of the business rules or model learning, and finally the fusion weight vector is obtained.

[0067] The fusion weight vector is ensured to be consistent in dimension with the structure data vector and the entity twin vector, wherein each element of the fusion weight vector corresponds to a weight value (value range 0-1) of a feature dimension, and for each feature dimension , the value of the th dimension of the structure data vector is multiplied by the weight of the th dimension of the fusion weight vector , the value of the th dimension of the entity twin vector is multiplied by , and the two products are added to obtain the value of the th dimension of the fusion feature vector. According to this rule, all dimensions are traversed to finally form a complete fusion feature vector. This vector dynamically integrates the static features of structured data and the multi-modal dynamic features of the entity twin vector through the weight allocation driven by business rules.

[0068] The fusion feature vector is arranged and stacked in order according to the unique identifier (such as a customer ID) of the entity as an independent row vector or column vector to form a two-dimensional matrix. If there are N entities, and each fusion feature vector has a dimension of D, then the entity feature matrix has a dimension of N x D. 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.

[0069] The essence of the embodiment is to determine the weight by parsing the business rule semantics through the generative large model, and to perform element-by-element weighted summation on the structured data vector and the entity twin vector, thereby constructing an entity feature matrix integrating structured attributes and multi-modal dynamic features, and realizing comprehensive digital modeling of the entity.

[0070] In step S20103, the generative large model is used to analyze the entity feature matrix to generate a candidate entity set.

[0071] In the embodiment, the entity feature matrix is input into the pre-trained generative large model LLaMA-2. The generative large model analyzes the distribution pattern, similarity and association rules of each entity feature vector in the matrix 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 to text descriptions such as "high consumption frequency + young customer group"), and then identifies dense areas or abnormal points in the feature space based on the entity generation logic implied by the business rules (such as customer segmentation and risk rating), to generate potential candidate entity initial names or initial category labels.

[0072] With the semantic understanding and feature association capability of the generative large model, the entity feature matrix integrating structured data and twin features is analyzed. Through clustering, association rule mining or semantic matching, etc., the entity combination or associated entity group with potential business value is identified. The candidate entity set may contain different name expressions (such as aliases, abbreviations, historical former names, etc. non-standardized names) of the same entity, because the generative large model has cross-language and cross-expression semantic alignment capability, and can identify that "Huayi" and "Huayi Technology Co., Ltd." and "HY" refer to the same entity in different forms. If the entity feature matrix contains multi-modal text data containing aliases, the model will aggregate different names into candidate entities based on semantic association. However, these candidate entities may not have been standardized and need to be standardized by the entity disambiguation module later.

[0073] In step S20104, a rule engine is introduced, and the business rules are loaded using the rule engine to filter and standardize the candidate entity set, to obtain a plurality of entities.

[0074] In this embodiment, first, the rule engine (such as Drools, EasyRule, etc.) is initialized, and the business rules of the target field (such as the 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 level", "abnormal behavior frequency less than or equal to 2 times in the past 30 days", etc.) are loaded into the engine in the form of JSON, XML or rule script. The engine traverses the candidate entity set generated by the generative large model, matches each candidate entity according to its feature description (such as "TOP 10% customers in East China region in terms of consumption frequency"), filters the entities that meet the rules through conditional filtering (such as numerical comparison, category matching, logical combination), and finally, based on the pre-set naming specification in the rules (such as "high-value customers-platinum level", "potential risk customers"), the standardized name of the filtered entities is calibrated to form an entity set that meets the business logic (such as an entity list containing "high net worth customers", "high-quality suppliers", etc.), ensuring the accuracy and business interpretability of the entity definition.

[0075] The standardization process refers to the whole process of loading business rules through a rule engine, performing rule matching, conditional filtering, and name calibration on the candidate entity set to eliminate the ambiguity, redundancy, and non-standardization of entity features, so as to meet the business rules and data governance standards of the target field.

[0076] Step S202, using the business data in the structured data to extract the direct association relationship between the entities, and associating the direct association relationship with the entity twin vector to obtain the explicit relationship.

[0077] In this embodiment, key business fields are extracted from multi-source data by a data analysis module, such as customer ID, transaction counterparty ID, transaction time and amount extracted from bank statements, parent company ID, subsidiary ID and holding ratio extracted from enterprise business information, etc. Using predefined field mapping rules (such as data dictionary or mapping table), the "customer ID", "counterparty ID" and other fields are accurately mapped to the unique entity ID in the entity feature matrix, and invalid ID records are removed through validity check. Based on industry business rules, a relationship type dictionary is constructed, such as defining "transaction counterparty" (applicable to fund transfer records), "guarantor" (applicable to credit contracts), "parent-subsidiary company" (applicable to enterprise associations with holding ratio > 50%) and other relationship types, and associating data features through conditional expressions (such as "contract type = procurement" and relationship type "supplier"). For each business data record, generate original triples in the form of "entity 1-relationship type-entity 2", such as transaction record (customer ID = 1001, counterparty ID = 1002, transaction type = transfer) corresponding to triple (1001, transaction counterparty, 1002), and de-duplicate through timestamp or business serial number, and aggregate repeated relationship records according to the latest time or cumulative value (such as cumulative transaction amount). When associating entity twin vectors, through distributed cache or batch query interface, the corresponding twin vector is retrieved from the entity feature matrix according to the entity ID, and the vectors of entity 1 and entity 2 are embedded in the relationship representation in JSON format to form explicit relationships.

[0078] In step S203, the entity twin vector corresponding to the entity is analyzed by the generative large model to obtain a potential relationship.

[0079] In step S203, the entity twin vector corresponding to the entity is analyzed by the generative large model to obtain a potential relationship.

[0080] In step S20301, the entity twin vector corresponding to the entity is analyzed by the generative large model to obtain a potential hypothesis relationship and confidence.

[0081] In this embodiment, the input entity twin vector is preprocessed in multiple stages: the feature dimension difference is eliminated by standardization technology, PCA or t-SNE algorithm is used to reduce high-dimensional vector to model adaptive dimension, and sliding window mechanism is used to frame time series vector 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 text proportion greater than 40%), the prompt word template is automatically selected (such as based on the semantic relationship between "keyword extraction results" and behavior characteristics "time series pattern summary", the potential association between entities is inferred), and a small amount of sample examples are inserted to guide the model to generate relationship hypotheses consistent with cross-domain logic.

[0082] In the model inference stage, the multi-head attention mechanism is disassembled into two layers of intra-modal and inter-modal processing: the first layer of attention focuses on single-modal features (such as the weight of noun phrases in text vectors, color distribution features in image vectors), extracting local key features; the second layer of attention establishes cross-modal correlation (such as the visual semantic alignment of the text "R&D investment" and the image "laboratory scene"), quantifying the feature interaction strength through an attention weight matrix. The model generates hypotheses based on a hybrid relationship knowledge base (integrating general common sense relationships and extensible industry rules), for example: when the "market share" value in the vector is close and the "product line overlap degree" is greater than 0.6, trigger the "competitive relationship" reasoning module, combined with the historical relationship distribution of similar feature combinations in the training data (such as 82% of the cases corresponding to competitive relationships), generate an initial relationship hypothesis and confidence.

[0083] In an optional implementation, the confidence can also be calibrated. Specifically, a Bayesian network model can be introduced, with multi-modal feature evidence (such as text evidence, behavior evidence, image evidence) as network nodes, and joint probability calculated through conditional probability table to make posterior correction to the confidence.

[0084] Step S20302, constructing a pre-defined rule base, using the confidence to screen the potential hypothesis relationship based on the pre-defined rule base, and obtaining a potential relationship.

[0085] In this embodiment, for target domains (such as financial risk control, customer portrait, etc.), business logic rules, industry standard rules, and data verification rules are integrated to form a structured rule set, where the business logic rules include confidence threshold for relationship establishment (such as "cross-industry association relationship requires confidence greater than 0.8"), feature matching conditions (such as "associated entities must have high-frequency transaction records"), industry standard rules cover domain-specific constraints (such as "parent-subsidiary relationship requires holding percentage greater than or equal to 50%", "partners must have industry access qualifications"), and data verification rules involve multi-modal feature consistency requirements (such as "potential relationship must match at least 2 or more modal features", "time series data must comply with business cycle rules"). All rules are stored in the rule base in structured forms such as JSON, XML, or rule scripts, forming a pre-defined rule base,

[0086] The potential hypothesis relationship and confidence output by the generative large model (such as "customer A and customer B have a supply chain relationship, with a confidence of 0.72") is input into the rule base, and the conditions in the rule base are matched one by one: for each potential relationship, check whether its confidence meets the preset threshold of the rule (such as 0.7), and verify whether the relationship description conforms to the business logic (such as "supply chain relationship" needs to exist high-frequency transaction record features in the behavior vector of the entity). Finally, only the relationships that meet both the confidence requirement and the rule condition are retained as potential relationships (such as "supply chain relationship" with a confidence of 0.75 and transaction records), and hypotheses that do not meet the rules are rejected (such as "cooperation relationship" with a confidence of 0.6), ensuring that the potential relationship has both data-driven credibility and business domain logic constraints.

[0087] Step S204, constructing a knowledge graph based on the entity, the explicit relationship and the potential relationship.

[0088] In this embodiment, the entities labeled by business rules are selected as nodes of the knowledge graph, and each node is attached with its corresponding entity twin vector (containing text semantics, behavior patterns, visual attributes and other multi-modal dynamic features). The explicit relationship extracted from the structured business data and associated with the entity twin vector, and the potential relationship obtained by the generative large model analysis and rule library screening are used as edges of the graph, and each edge is labeled with relationship type (such as "transaction", "supply chain"), confidence (only for potential relationship) and associated entity dynamic features. According to the association logic between entities, the nodes and edges are structured and connected to form a knowledge graph containing entity nodes, explicit edges, potential edges and multi-dimensional attributes. The graph supports real-time evolution with the update of entity twin vectors, and can dynamically reflect the temporal changes of entities and their associated relationships, providing structured knowledge representation and multi-dimensional reasoning basis for customer profile analysis, business association mining and other scenarios.

[0089] The knowledge graph takes entity twin vectors and time-sensitive edge relationships as the core, and covers the whole process dynamics of entity feature evolution, relationship weight update and graph structure adjustment.

[0090] Step S3, the customer profile includes an interaction log, and the edge relationship weight of the knowledge graph is calculated using the interaction log.

[0091] The calculation of the edge relationship weight of the knowledge graph using the interaction log includes the following sub-steps:

[0092] Step S301, adding time information in the interaction log to the knowledge graph to form a time-series knowledge graph.

[0093] In the present embodiment, first, the timestamp field and the corresponding entity interaction event (such as "customer A and customer B transact on 2025-06-01") are extracted from the interaction log (such as customer behavior log, transaction record log, etc.), the event is associated to the entity node in the knowledge graph through the entity ID mapping rule, then, the time attribute is added to the explicit relationship and the potential relationship in the graph, for example, the fields of "relationship establishment time" and "latest update time" are recorded in the relationship edge, for the periodic or continuous relationship (such as long-term cooperation agreement), the time interval (start time-end time) is marked, finally, the time dimension is introduced into the knowledge graph structure, the entity state transition chain with timestamp (such as the node sequence of the credit rating of entity A changing over time) is constructed, the time sequence knowledge graph with time as the main axis and entity and relationship as the node and edge is formed, the functions of time slicing query and time sequence correlation analysis are supported, and the whole cycle tracking of the evolution process of entity and its relationship is realized.

[0094] The time sequence knowledge graph is extended in the time dimension of the knowledge graph, the graph structure with timestamp is constructed through the time information of the interaction log, and the relationship analysis according to time slicing is supported.

[0095] In step S302, time sequence data of edge relationship in the time sequence knowledge graph is extracted to form a time sequence vector.

[0096] In the present embodiment, all edges (including explicit relationship and potential relationship) in the time sequence knowledge graph are traversed, the time attribute (such as the timestamp of a single interaction record, the time interval of a continuous relationship) of each edge is extracted, the time sequence data is extracted according to the entity pair (such as "entity A-entity B"), for the instantaneous edge relationship (such as a single transaction), the numerical representation (such as Unix timestamp or converted to numerical coding of "year-month-day") of the timestamp is extracted and arranged in time sequence to form a one-dimensional time sequence, for the continuous edge relationship (such as long-term cooperation agreement), the relationship duration (such as the number of days of end time-start time) and the time interval start point are calculated and converted to numerical vector. Finally, all time sequence data is standardized (such as normalized to [0, 1] interval or Z-score standardization), and the time sequence vector is formed by splicing according to the entity pair dimension, which captures the time distribution law and evolution characteristics of the relationship between entities.

[0097] In step S303, a weight calculation model is constructed based on a time sequence graph neural network, and an edge relationship weight is obtained by using the weight calculation model according to the time sequence vector and the semantic vector.

[0098] In this embodiment, the historical customer files before a predetermined time (such as two years) are used, the semantic vector is extracted from the historical text data by using a large model, the behavior vector is extracted from the behavior data, the visual vector is extracted from the image data, and the behavior and visual vectors are converted into historical description text. The historical entity twin vector is generated by fusing through a dynamic gating mechanism. Secondly, the timestamp is extracted from the interaction log two years ago, the historical time sequence data is grouped according to the entity pair to form the historical time sequence vector, and the historical time vector is extracted from the historical time sequence vector to obtain the 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 law in the historical structured data, the explicit relationship weight (such as "historical transaction amount ≥ 1 million" marked as 1.0) is labeled, the dynamic features of the historical entity twin vector are analyzed by the generative large model to obtain the potential hypothesis relationship and confidence, and after screening combined with the historical pre-defined rule library, the potential relationship label (0-1 numerical value) is obtained. Finally, the "historical time sequence features + historical semantic features" are used as input features, the historical edge weight label is used as a supervision signal, the training sample is constructed, the weight calculation model is trained by using the training sample, and the pre-trained weight calculation model is obtained. Finally, the time sequence vector and the semantic vector are input into the weight calculation model to obtain the edge relationship weight.

[0099] 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 time sequence importance of the historical interaction data can be weighted, the influence of long-term interaction can be suppressed, the weight of recent interaction can be strengthened, the graph can dynamically reflect the business changes, and by setting a self-attention layer after the input layer, the association weight of the time sequence vector and the semantic vector can be explicitly calculated, the black box problem of the original model relying on neural network implicit learning features can be solved, the edge relationship weight is more explainable and conforms to the business logic, and the mining ability of the knowledge graph for implicit relationships is improved.

[0100] The weight calculation model based on the time sequence graph neural network comprises: a time decay module in the input layer of the weight calculation model; a self-attention layer after the input layer; and the weight calculation model is constructed based on the time decay module and the self-attention layer.

[0101] The time decay module satisfies the following formula:

[0102]

[0103]

[0104] wherein, is an entity and an entity a decayed time sequence vector, This represents the total number of interactions between entities. For the first The time decay factor of each interaction For the first The original time series vector of the second interaction, It is a natural exponential function. The empirical attenuation coefficient is... For the first The time interval between the current interaction and the current time.

[0105] The self-attention layer satisfies the following formula:

[0106]

[0107]

[0108] in, For entities and entity The fused feature vector, For self-attention weights, For entities and entity The vector obtained by linear projection of the decayed time series vector. For entities and entity The semantic vector after linear projection, is the dimension of the projected feature vector.

[0109] The above empirical attenuation coefficient The value is determined through grid search, with a preset empirical attenuation coefficient. The candidate value set is used to calculate the Spearman correlation coefficient between the edge weights and the actual business intensity on historical data, and the value with the highest correlation coefficient is selected. Value, such as in financial risk control scenarios At that time, the recent decline in the weight of transactions is more in line with the law of risk transmission.

[0110] Step S4: Adjust the knowledge graph based on the edge relationship weights.

[0111] The adjustment of the knowledge graph based on the edge relation weights specifically includes the following sub-steps:

[0112] Step S401: Set a weight threshold and compare the weight threshold with the edge relationship weight;

[0113] In the embodiment, a weight threshold is set according to business requirements or field rules (such as a correlation strength threshold in risk assessment, a credibility threshold of a 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.

[0114] In step S402, the edge corresponding to the edge relationship weight in the knowledge graph is updated or deleted according to the comparison result.

[0115] In the embodiment, for the edge with a weight value greater than or equal to the threshold, it is determined as a strong correlation relationship and retained (such as marked as "highly credible association"), and for the edge with a weight value less than the threshold, it is determined as a weak correlation relationship and deleted.

[0116] As shown in Figure 2 In addition, the present application also provides a customer profile knowledge graph construction system based on a large model driver, which comprises a processor, an input device, an output device and a memory, wherein the processor, the input device, the output device and the memory are connected to each other, the memory is used to store a computer program, the computer program comprises program instructions, the processor is configured to call the program instructions, and the related steps of the related embodiments of the customer profile knowledge graph construction method based on the large model driver are executed.

[0117] The customer profile knowledge graph construction system based on the large model driver provided by the present application can integrate each functional component in one processing component, or each component can exist physically alone, or two or more components can be integrated in one component. The integrated components can be realized in the form of hardware or in the form of software functions.

[0118] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the above embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and the description of the present application.

Claims

1. A large model driving-based customer profile knowledge graph construction method, characterized in that, The method comprises: encoding the customer profile using large model technology, and fusing the results of the encoding to generate entity twin vectors, comprising: the customer profile includes text data, behavior data and image data; using large model technology to respectively extract semantic features from the text data, behavior features from the behavior data, and visual features from the image data, to obtain semantic vectors, behavior vectors and visual vectors; respectively converting the semantic vectors, behavior vectors and visual vectors into semantic description text, behavior description text and visual description text; using a dynamic gating mechanism to fuse the semantic vectors, behavior vectors, visual vectors, semantic description text, behavior description text and visual description text to obtain entity twin vectors; the entity twin vectors satisfy the following formula: wherein, is a function of time updated entity twin vector, is a Sigmoid function, is a pre-trained generative large language model, is a description text of the th modality at time is a feature vector of the th modality at time is a feature vector of the th modality at time using the structured data in the customer profile and the entity twin vectors to construct a knowledge graph; the customer profile includes interaction logs, and the edge relationship weights of the knowledge graph are calculated using the interaction logs; adjusting the knowledge graph based on the edge relationship weights.

2. The customer archive knowledge graph construction method based on a large model driver according to claim 1, characterized in that, the use of structured data in the customer profile and the entity twin vectors to construct a knowledge graph comprises: extracting business rules based on the business needs of the target field, and obtaining entities according to the structured data of the customer profile and the business rules; extracting direct association relationships between the entities using business data in the structured data, and associating the direct association relationships with the entity twin vectors to obtain explicit relationships; analyze the entity twin vectors corresponding to the entities using a generative large model to obtain potential relationships; construct a knowledge graph based on the entities, explicit relationships and potential relationships. 3.The method of claim 2, wherein, the entity obtained according to the structured data of the customer profile and the business rules comprises: vectorizing the structured data of the customer profile to obtain a structure data vector; weighting and fusing the structure data vector and the entity twin vector to construct an entity feature matrix; using a generative large model to analyze the entity feature matrix to generate a candidate entity set; introducing a rule engine, loading the business rules using the rule engine, and filtering and standardizing the candidate entity set to obtain multiple entities.

4. The customer archive knowledge graph construction method based on a large model driver according to claim 3, characterized in that, The structured data of the customer profile includes numerical structure data and category structure data, and the vectorization of the structured data of the customer profile to obtain a structure data vector comprises: normalizing the numerical structure data to generate a numerical feature vector; pre-training embedding coding is performed on the category structure data to generate a category feature vector; splicing the numerical feature vector and the category feature vector to obtain a structure data vector.

5. The customer archive knowledge graph construction method based on a large model driver according to claim 2, characterized in that, the analysis of the entity twin vectors corresponding to the entities by a generative large model to obtain potential relationships comprises: using a generative large model to analyze the entity twin vectors corresponding to the entities to obtain potential hypothesis relationships and confidence levels; A predefined rule library is constructed, and the potential hypothesis relationships are screened based on the predefined rule library and the confidence to obtain potential relationships.

6. The customer archive knowledge graph construction method based on a large model driver according to claim 1, characterized in that, The calculating of the edge relationship weight of the knowledge graph based on the interaction log comprises: Adding time information in the interaction log into the knowledge graph to form a time sequence knowledge graph; Extracting time sequence data of the edge relationship in the time sequence knowledge graph to form a time sequence vector; A weight calculation model is constructed based on a time sequence graph neural network, and the edge relationship weight is obtained based on the time sequence vector and the semantic vector by using the weight calculation model.

7. The customer archive knowledge graph construction method based on a large model driver according to claim 1, characterized in that, The adjusting of the knowledge graph based on the edge relationship weight comprises: Setting a weight threshold, comparing the weight threshold with the edge relationship weight; According to the comparison result, the edge corresponding to the edge relationship weight in the knowledge graph is updated or deleted.

8. A large model driving-based customer profile knowledge graph construction system, characterized in that, Comprise: A processor, an input device, an output device and a memory, the processor, the input device, the output device and the memory are connected with each other, wherein the memory is used for storing a computer program, the computer program comprises program instructions, the processor is configured to call the program instructions, and executes the method for constructing a customer profile knowledge graph based on a large model driving according to any one of claims 1 to 7.

Images