Customer twin construction method based on multi-source data fusion and large model driving

Abstract

The application provides a customer twin construction method based on multi-source data fusion and large model driving, and belongs to the technical field of customer relationship management. It comprises: collecting structured data, unstructured data and semi-structured data; generating a global fusion model; extracting customer basic attribute labels, constructing a customer knowledge graph based on an entity-relation-entity triple structure, and storing "customer-product" and "customer-service" interaction relationships; inputting customer behavior sequence data into a time sequence fusion transformer model, capturing long-term dependence and short-term fluctuation characteristics using the self-attention mechanism in the time sequence fusion transformer model, identifying behavior patterns, and outputting a dynamic feature vector containing time evolution information; and weighting and fusing the customer basic attribute labels and the dynamic feature vector to construct a customer digital twin. The completeness, timeliness and prediction accuracy of the customer portrait are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of customer relationship management, and particularly relates to a customer twin construction method based on multi-source data fusion and large model driving. BACKGROUND

[0002] With the deepening of enterprise digital transformation, customer data sources are increasingly diversified and widely distributed in CRM systems, e-commerce platforms, social media and Internet of Things devices, showing heterogeneous characteristics of coexistence of structured, semi-structured and unstructured data.

[0003] Traditional customer portrait technology mostly relies on a single data source, uses a static label system and a rule engine for modeling, and is difficult to realize effective integration and semantic unification of cross-domain data, resulting in fragmentation of customer views and inability to fully reflect their real preferences and behavior logic. SUMMARY

[0004] To solve at least one aspect of the technical problems in the background art, the application provides a customer twin construction method based on multi-source data fusion and large model driving, which realizes dynamic modeling and real-time updating of customer full-dimensional features through multi-source data fusion and large model driving, and improves the completeness, timeliness and prediction accuracy of customer portrait.

[0005] The technical scheme adopted by the application is as follows:

[0006] The first aspect of the application provides a customer twin construction method based on multi-source data fusion and large model driving, comprising:

[0007] The RESTful API interface is used to respectively connect enterprise internal systems, external platforms and Internet of Things devices to collect structured data, unstructured data and semi-structured data in real time;

[0008] A feature extraction model is deployed locally in each data source, and a federated learning framework is used for distributed training, only the local model parameters are encrypted and uploaded to the center node, the center node performs aggregation operation on multiple local model parameters to generate a global fusion model;

[0009] Customer basic attribute labels are extracted from the cleaned fusion data, including demographic information, consumption ability level and product preference category, a customer knowledge graph is constructed based on entity-relation-entity triple structure, and "customer-product" and "customer-service" interaction relationships are stored;

[0010] The customer behavior sequence data is input into a time sequence fusion transformer model, long-term dependence and short-term fluctuation characteristics are captured by using a self-attention mechanism in the time sequence fusion transformer model, behavior patterns are identified, and a dynamic feature vector containing time evolution information is output, which is used to represent a current interest state and a future behavior trend of a customer.

[0011] The customer basic attribute label is weighted and fused with the dynamic feature vector to construct a customer digital twin.

[0012] According to an embodiment of the present application, the structured data includes customer basic information and order records, the unstructured data includes user comments and customer service audio-to-text content, and the semi-structured data includes user behavior logs and click stream data.

[0013] According to an embodiment of the present application, the feature extraction model is deployed locally at each data source, distributed training is performed using a federated learning framework, only local model parameters are encrypted and uploaded to a central node, the central node performs aggregation operations on a plurality of local model parameters to generate a global fusion model, and the operations are specifically as follows:

[0014] A local feature extraction model is initialized at each data source, the local feature extraction model includes a multi-layer perceptron subnetwork for processing structured data, a lightweight BERT embedding layer for processing unstructured text data, and a projection layer for aligning output dimensions;

[0015] Each data source uses local data to locally train the local feature extraction model and update model parameters, and original data is not uploaded during the training process;

[0016] The updated local model parameters are encrypted and uploaded to the central node and a federated averaging algorithm or a momentum-based aggregation algorithm is performed to generate an updated global fusion model;

[0017] The global fusion model parameters are distributed to each local node for the next round of training iteration until the model converges;

[0018] Based on the converged global fusion model, cross-domain feature extraction and semantic alignment are performed on new samples.

[0019] According to an embodiment of the present application, the customer basic attribute label is extracted from the cleaned fusion data, and the operations are specifically as follows:

[0020] Customer static attribute fields including age, gender, region, registration duration, total annual consumption, order frequency, and membership level are extracted from the fusion data;

[0021] The static attribute fields are classified and labeled based on a preset rule or a clustering algorithm to generate a basic label.

[0022] Based on the basic label, a customer and an entity associated therewith are organized into a triple in the form of "subject-predicate-object" ;

[0023] The triple is imported into a graph database to construct a customer knowledge graph.

[0024] According to an embodiment of the present application, the customer behavior sequence data is input into a time fusion transformer model, and a self-attention mechanism in the time fusion transformer model is used to capture long-term dependence and short-term fluctuation characteristics and identify behavior patterns, specifically:

[0025] The behavior sequence of the customer within a preset time window is extracted from the fusion data, including page browsing, product adding, order payment, customer service consultation, content liking, and search keywords, and is sorted by timestamp and constructed into multivariate time series data;

[0026] The time series data is input into a time fusion transformer model, and a gating mechanism is used to separate static features and dynamic features, and a self-attention layer is used to calculate the dependence weight between different time steps;

[0027] An interpretable attention weight analysis is performed to identify behavior patterns.

[0028] According to an embodiment of the present application, the dynamic feature vector containing time evolution information is output, which is used to represent the current interest state and future behavior trend of the customer, specifically:

[0029] A high-dimensional hidden state vector sequence is output by the time fusion transformer model, and each time step corresponds to a dynamic feature vector, representing the interest intensity and behavior tendency of the customer at that moment;

[0030] The hidden state vector at the final time step is nonlinearly mapped to generate a current interest state vector, which is used to depict the focus of the customer at the current moment;

[0031] The prediction head of the time fusion transformer model is used to output the probability distribution of multiple future time steps to generate a future behavior trend prediction result.

[0032] According to an embodiment of the present application, the customer basic attribute label and the dynamic feature vector are fused by weighting to construct a customer digital twin, specifically:

[0033] The customer basic attribute label is encoded into a fixed-dimensional static feature vector, and a one-hot encoding or embedding representation method is used to map it to the same high-dimensional semantic space as the dynamic feature vector;

[0034] An active degree index of the customer is obtained, which is calculated based on the frequency of behavior events in the past 30 days, and the formula is A=log(1+Nevents ), wherein A represents an activity score, N events is the total number of valid behaviors of the customer in the statistical period;

[0035] The fusion weight is dynamically calculated according to the activity score A, wherein the dynamic feature weight calculation formula is w d = sigma (k•(A-A0)), wherein sigma is a Sigmoid function, k is a gain coefficient, A0 is a reference activity threshold, and the static feature weight calculation formula is w s = 1-w d .

[0036] The static feature vector v s and the dynamic feature vector v d are weighted and fused to generate a customer comprehensive feature vector v twin , wherein v twin = w s •v s + w d •v d .

[0037] According to one embodiment of the present application, it further comprises:

[0038] The constructed customer digital twin is input into a large language model, the customer state and trend are converted into a natural language description through prompt word engineering, an individualized service strategy or risk warning suggestion is generated, and the output is output to a CRM system, an intelligent customer service platform or a marketing automation system through an API interface.

[0039] The second aspect embodiment of the present application provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the program to realize the customer twin construction method based on multi-source data fusion and large model driving in any embodiment of the first aspect.

[0040] The third aspect embodiment of the present application provides a non-volatile computer storage medium, which stores computer executable instructions, and the computer executable instructions are executed by a processor to realize the customer twin construction method based on multi-source data fusion and large model driving in any embodiment of the first aspect.

[0041] Due to the adoption of the above technical solutions, the present application has the following beneficial effects:

[0042] According to the method for constructing a customer twin based on multi-source data fusion and large model driving provided in the first aspect of the application, first, the structured data (such as user information and order records), unstructured data (such as user comments and customer service audio text conversion) and semi-structured data (such as behavior logs and click streams) are collected in real time through the RESTful API interface to the enterprise internal system (such as CRM and ERP), external platform (such as e-commerce platform and social media) and Internet of Things device (such as intelligent terminal), solving the problem of serious data island and lagging update in traditional systems and ensuring the comprehensiveness and timeliness of customer data; second, the feature extraction model is deployed locally in each data source and distributed training is performed using the federated learning framework, and only the encrypted local model parameters are uploaded to the center node to generate a global fusion model, which not only realizes collaborative modeling of cross-domain data but also effectively protects the original data from being out of the domain, meets the increasingly stringent privacy compliance requirements and improves the security and scalability of the system; then, the customer basic attribute label is extracted from the cleaned fusion data, and the customer knowledge graph is constructed based on the “subject-predicate-object” triple structure, and the complex interaction relationship between the customer and the product and service is stored in the form of a graph structure, which not only enhances the semantic expression ability of the customer static portrait but also supports deep relationship reasoning and correlation analysis, providing support for accurate recommendation and risk identification; further, the customer behavior sequence data is input into the timing fusion transformer model, and the long-term dependence and short-term fluctuation characteristics between behaviors are captured using the self-attention mechanism, which can accurately identify the key behavior path such as “browsing→adding to cart→purchasing” and its time evolution law, output a dynamic feature vector containing interest state evolution and future trend prediction, and significantly improve the prediction ability of the customer's short-term intention and long-term preference; finally, the static basic attribute label and dynamic behavior feature vector are weighted and fused to construct a customer digital twin that integrates the state layer, evolution layer and interaction layer, and the fusion mechanism can adaptively adjust the weight according to the customer activity to ensure that the real-time behavior of high-activity customers is fully responded to, while the long-term portrait stability of low-activity customers is retained, thereby forming a comprehensive and sensitive customer virtual mapping. In summary, the method combines multi-source heterogeneous data fusion, privacy-protected collaborative training, knowledge graph enhancement and large model timing modeling, overcomes the defects of traditional customer portrait such as static, fragmentation and weak prediction ability, and significantly improves the depth of customer understanding and the accuracy of intelligent decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0043] The drawings described herein are intended to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their description serve to explain the present application. They are not intended to be an improper limitation on the present application. In the drawings:

[0044] Figure 1A flowchart of a customer twin construction method based on multi-source data fusion and large model driving is provided for an embodiment of the present application.

[0045] Figure 2 A structural schematic diagram of an electronic device is provided for an embodiment of the present application.

[0046] Reference signs:

[0047] 810, processor; 820, communication interface; 830, memory; 840, communication bus. DETAILED DESCRIPTION

[0048] In order to more clearly illustrate the overall concept of the present application, the following will be described in detail with reference to the accompanying drawings.

[0049] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and therefore the scope of protection of the present application is not limited by the specific embodiments disclosed below. It should be noted that the embodiments of the present application and the features in each embodiment can be combined with each other without conflict.

[0050] In the present application, unless otherwise explicitly specified and limited, the first feature is "on" or "under" the second feature, which can be direct contact between the first and second features, or indirect contact between the first and second features through an intermediate medium. In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0051] As Figure 1 shown, the first aspect embodiment of the present application provides a customer twin construction method based on multi-source data fusion and large model driving, comprising:

[0052] Step 100, through the RESTful API interface, respectively connecting the enterprise internal system, the external platform and the Internet of Things device, real-time collection of structured data, unstructured data and semi-structured data.

[0053] Step 200, locally deploying a feature extraction model for each data source, using a federated learning framework for distributed training, and only uploading the local model parameters to the center node after encryption, and the center node performs aggregation operation on multiple local model parameters to generate a global fusion model.

[0054] Step 300, extracting customer basic attribute tags from the cleaned fusion data, the customer basic attribute tags including demographic information, consumption ability level, and product preference category, constructing a customer knowledge graph based on entity-relation-entity triple structure, and storing "customer-product" and "customer-service" interaction relationships.

[0055] Step 400, inputting customer behavior sequence data into a time series fusion transformer model, capturing long-term dependencies and short-term fluctuations using self-attention mechanisms in the time series fusion transformer model, identifying behavior patterns, and outputting dynamic feature vectors containing time evolution information to represent current interest states and future behavior trends of customers.

[0056] Step 500, weighting and fusing customer basic attribute tags and dynamic feature vectors to construct a customer digital twin.

[0057] In step 100, the RESTful API interface is used to connect with enterprise internal systems (such as CRM, ERP), external platforms (such as e-commerce platforms, social media), and Internet of Things devices (such as smart terminal interaction devices) to collect structured data (such as user basic information, order records), unstructured data (such as user comments, customer service audio transcription), and semi-structured data (such as click stream, behavior logs) in real time. This step realizes the unified access and continuous update of multi-source heterogeneous data, breaks the problem of incomplete customer view caused by data silos in traditional systems, ensures the comprehensiveness, real-time nature, and system compatibility of data collection, and lays a data foundation for subsequent high-precision customer modeling.

[0058] In step 200, feature extraction models are deployed locally in each data source, and federated learning framework is used for distributed training. Only the local model parameters are encrypted and uploaded to the center node, and the center node performs aggregation operations (such as FedAvg algorithm) on multiple local model parameters to generate a global fusion model. This step completes cross-domain collaborative modeling without centralized raw data, effectively solves the data privacy and compliance risks, and realizes a safe fusion mechanism of "data not moving, model moving"; at the same time, by introducing a lightweight BERT embedding layer and a multi-layer perception subnetwork, unified semantic encoding of text and numerical features is realized, the consistency of heterogeneous data representation is improved, and the generalization ability of the model is enhanced.

[0059] In step 300, customer base attribute labels are extracted from the cleaned fusion data, including demographic information (such as age, gender, region), consumption level (such as high / medium / low consumption groups), and product preference categories (such as smart home, baby products), and a customer knowledge graph is constructed based on the "entity-relation-entity" triple structure, storing "customer-product" purchase relationships, "customer-service" consultation relationships, and other interaction information. This step converts static customer attributes into a structured semantic network, not only improving the explainability of customer portraits, but also supporting graph-based association reasoning (such as "customers who purchase product A also tend to purchase product B"), providing deep insights for personalized recommendations, cross-selling, and other scenarios.

[0060] In step 400, customer behavior sequence data (such as timestamped event streams such as page views, adds to cart, payments, consultations) is input into a temporal fusion transformer (TFT) model, which uses its self-attention mechanism to capture long-term dependencies (such as seasonal purchasing patterns) and short-term fluctuations (such as increased clicks during promotions) between different time steps, identify key behavior patterns (such as "the probability of purchasing high-end mobile phones increases significantly within 72 hours after browsing"), and output dynamic feature vectors containing time evolution information, which are used to represent the current interest state and future behavior trends of customers. This step breaks through the limitations of traditional statistical models in modeling temporal dependencies, enabling accurate perception and trend prediction of customer intentions, significantly improving the accuracy and timeliness of customer response prediction.

[0061] In step 500, customer base attribute labels are encoded into static feature vectors, and the dynamic feature vectors output in step 400 are weighted and fused to construct a customer digital twin. The fusion weights are dynamically adjusted according to customer activity (such as giving higher dynamic weights to high-activity customers), ensuring that the twin can reflect both long-term stable attributes and respond sensitively to recent behavior changes. The final customer digital twin contains a state layer (current attributes), an evolution layer (future trends), and an interaction layer (touchpoint history), and is updated at a minute level through an incremental learning mechanism. This step realizes the organic integration of static and dynamic features, upgrading the customer portrait from a "static snapshot" to a "dynamic living body", significantly improving the decision support capability in precise marketing, intelligent customer service, and churn warning business scenarios.

[0062] According to the method for constructing a customer twin based on multi-source data fusion and large model driving provided in the first aspect of the application, first, the structured data (such as user information and order records), unstructured data (such as user comments and customer service audio-to-text content), and semi-structured data (such as behavior logs and click streams) are collected in real time through the RESTful API interface to interface enterprise internal systems (such as CRM and ERP), external platforms (such as e-commerce platforms and social media), and Internet of Things devices (such as smart terminals), solving the problem of serious data silos and lagging updates in traditional systems and ensuring the comprehensiveness and timeliness of customer data. Second, feature extraction models are deployed locally in each data source and distributed training is performed using a federated learning framework. Only encrypted local model parameters are uploaded to the central node to generate a global fusion model, which not only enables collaborative modeling of cross-domain data but also effectively protects the original data from being out of the domain, meets increasingly stringent privacy compliance requirements, and improves the security and scalability of the system. Third, customer basic attribute labels are extracted from the cleaned fusion data, and a customer knowledge graph is constructed based on the "subject-predicate-object" triple structure, storing the complex interaction between customers and products and services in the form of a graph structure, which not only enhances the semantic expression ability of the customer static portrait but also supports deep relationship reasoning and correlation analysis, providing support for accurate recommendation and risk identification. Fourth, customer behavior sequence data is input into a time series fusion transformer model, which uses its self-attention mechanism to capture long-term dependencies and short-term fluctuation characteristics between behaviors, accurately identifies key behavior paths such as "browsing -> adding to cart -> purchasing" and their time evolution rules, and outputs dynamic feature vectors containing interest state evolution and future trend prediction, significantly improving the prediction ability of customer short-term intentions and long-term preferences. Finally, the static basic attribute labels and dynamic behavior feature vectors are weighted and fused to construct a customer digital twin that integrates state, evolution, and interaction layers. This fusion mechanism can adaptively adjust the weights based on customer activity levels to ensure that the real-time behavior of highly active customers is fully responded to while maintaining the stability of the long-term portrait of less active customers, thereby forming a comprehensive and sensitive customer virtual mapping. In summary, the method combines multi-source heterogeneous data fusion, privacy-protected collaborative training, knowledge graph enhancement, and large model time series modeling to overcome the limitations of traditional customer portraits, such as static, fragmented, and weak prediction ability, significantly improving the depth of customer understanding and the accuracy of intelligent decision-making.

[0063] In some embodiments of the application, the structured data includes customer basic information and order records, the unstructured data includes user comments and customer service audio-to-text content, and the semi-structured data includes user behavior logs and click stream data.

[0064] Structured data refers to data that can be easily stored in traditional row-column format, such as database tables. In this context, it mainly includes customer basic information (such as name, age, gender, contact information, etc.) and order records (such as product names purchased, purchase time, payment amount, etc.).

[0065] Structured data provides clear identifiers and numerical values, which can be used to quickly and accurately identify customers and classify them. This type of data is easy to query and analyze, supporting complex data mining through tools such as SQL, providing a solid foundation for market segmentation and trend analysis. Based on order records and other structured data, businesses can better understand customer consumption habits and preferences, thereby developing more personalized marketing strategies and service plans.

[0066] Unstructured data refers to data that does not have a predefined data model or is not organized in a certain predefined way, such as text files, images, audio, video, etc. Here, we specifically refer to user reviews and customer service transcript content.

[0067] Through the analysis of user reviews and customer service conversations, businesses can obtain direct feedback on product usage experience, service satisfaction, etc., which helps to discover potential problems and improve products or services. Using natural language processing technology, the sentiment of customers can be extracted from unstructured data, helping brands understand public opinion about them in order to adjust public relations strategies in a timely manner. Combined with the information provided by unstructured data, businesses can provide more personalized services, such as recommending more suitable products or solutions based on customers' specific needs.

[0068] Semi-structured data has a certain organizational form but does not fully comply with the traditional table structure of relational databases, such as XML, JSON format documents. Here we mainly refer to user behavior logs and clickstream data.

[0069] By collecting users' online behavior tracks (such as browsing history, search keywords, dwell time, etc.), businesses can track users' behavior paths and identify common behavior patterns. Based on the analysis of clickstream data, websites and applications can optimize their interface design and interaction processes, improving user access efficiency and satisfaction. With the help of dynamic information provided by semi-structured data, systems can respond to customer needs in real time and provide highly personalized recommended content, enhancing customer stickiness and loyalty.

[0070] In some embodiments of the present application, feature extraction models are deployed locally at each data source, and distributed training is performed using a federated learning framework. Only local model parameters are encrypted and uploaded to the central node, and the central node performs aggregation operations on multiple local model parameters to generate a global fusion model, specifically:

[0071] A local feature extraction model is initialized at each data source side, which includes a multi-layer perceptron subnetwork for processing structured data, a lightweight BERT embedding layer for processing unstructured text data, and a projection layer for aligning output dimensions;

[0072] Each data source locally trains the local feature extraction model using local data, updates the model parameters, and does not upload raw data during the training process;

[0073] The updated local model parameters are encrypted and uploaded to the central node and the federated averaging algorithm or momentum-based aggregation algorithm is executed to generate an updated global fusion model;

[0074] The global fusion model parameters are distributed to each local node for the next round of training iteration until the model converges;

[0075] Based on the converged global fusion model, cross-domain feature extraction and semantic alignment are performed on new samples.

[0076] A local feature extraction model is initialized at each data source side, containing a multi-layer perceptron subnetwork, a lightweight BERT embedding layer, and a projection layer. To address the heterogeneity of different data types, the local model adopts a modular design:

[0077] Multi-layer perceptron (MLP) subnetwork: used to process structured data (such as customer age, order amount) and extract high-order combinations of numerical features through nonlinear transformation;

[0078] Lightweight BERT embedding layer: performs semantic encoding on unstructured text data (such as user reviews, customer service conversations) to generate context-related vector representations;

[0079] Projection layer: maps outputs of different modalities to a unified dimensional vector space, ensuring that feature representations from each data source can be aligned and fused in subsequent stages.

[0080] This achieves localized adaptive modeling of multi-source heterogeneous data, avoiding the processing difficulties caused by inconsistent raw data formats.

[0081] Each data source uses local data for local training to update model parameters, and does not upload raw data during the training process. Each participant only iteratively optimizes the local model using their own data locally, without the need to share sensitive customer information (such as order details, behavior logs).

[0082] This mechanism adheres to the "data does not leave the domain" principle, complying with regulations such as the "Data Security Law" and the "Personal Information Protection Law", significantly reducing the risk of data leakage.

[0083] The updated local model parameters are encrypted and uploaded to the center node, and a federated average (FedAvg) or momentum-based aggregation algorithm is executed. The local model parameters are protected by homomorphic encryption or differential privacy perturbation technology before uploading to prevent the center node or third party from inferring the original data through the parameters.

[0084] The center node uses a federated average algorithm (weighted average of each node parameter) or a more advanced momentum aggregation algorithm (such as FedAdam) to integrate the knowledge of multiple local models and generate an updated global fusion model.

[0085] The aggregation process focuses on the co-evolution of model knowledge rather than data centralization, realizing a new collaboration paradigm of "model flow, data isolation".

[0086] Global model parameters are distributed to each local node for the next round of training until the model converges, forming a typical multi-round iteration federated learning process: in each round of communication, the local model is initialized based on the global model, trained locally, and then updated and uploaded, and the center aggregates to generate a new global model.

[0087] As the iteration proceeds, each local model gradually converges to a unified feature extractor with cross-domain generalization capability.

[0088] Based on the converged global fusion model, cross-domain feature extraction and semantic alignment are performed on new samples. The converged global model can be deployed on each node or the center system to encode new customer data uniformly and generate a semantic consistent feature vector.

[0089] For example, "high-consumption users" from an e-commerce platform and "VIP customers" from a CRM system can be identified as similar entities in the same vector space, achieving cross-system semantic alignment.

[0090] Through the mechanism of local modeling + encrypted parameter aggregation + multi-round federated iteration, a secure, efficient, and intelligent cross-domain feature fusion system is constructed. Not only does it protect the data sovereignty and privacy of each participant, but it also realizes knowledge collaboration and semantic unification of multi-source heterogeneous data, providing a high-quality and consistent feature basis for subsequent customer knowledge graph construction, dynamic behavior modeling, and digital twin generation.

[0091] In some embodiments of the present application, customer basic attribute labels are extracted from the cleaned fusion data, specifically:

[0092] The customer static attribute fields are extracted from the fusion data, including age, gender, region, registration duration, total annual consumption, order frequency, and membership level.

[0093] The static attribute fields are classified and labeled based on pre-set rules or clustering algorithms to generate basic labels:

[0094] Based on the basic label, the customer and the entity associated therewith are organized into triples in the form of "subject-predicate-object";

[0095] The triples are imported into a graph database to build a customer knowledge graph.

[0096] In the construction process of the customer digital twin, the customer basic attribute labels are extracted from the cleaned fused data and the knowledge graph is constructed. First, the customer static attribute fields are extracted from the unified data set that has completed privacy protection fusion and data cleaning, including age, gender, region, registration duration, annual total consumption, order frequency, and member level, and other characteristic dimensions that are long-term stable or change slowly. These fields reflect the inherent attributes and historical behavior accumulation of the customer, and are the basis for depicting the customer portrait. Subsequently, based on preset business rules (such as "annual total consumption ≥ 80,000 yuan" marked as "high consumption ability") or using unsupervised clustering algorithms (such as K-means, DBSCAN), the customers are automatically classified and labeled, generating a structured basic label system covering multiple dimensions such as demographic information, consumption ability level, and product preference category. On this basis, the customer and the entity associated therewith (such as product, service, and activity) are semantically organized in the form of triples, such as "Customer A - buys - smart speaker" and "Customer B - consults - after-sales service", realizing the structured expression of customer behavior and relationship. Finally, all triples are imported into a graph database (such as Neo4j or JanusGraph) in batches to build a customer knowledge graph, supporting efficient graph traversal, relationship reasoning, and path analysis.

[0097] The customer basic labels are generated by combining rules and algorithms, which not only ensures the interpretability of the labels, but also improves the objectivity and automation level of classification, avoiding the subjectivity and inefficiency of manual annotation. Second, the customer attributes and interaction relationships are converted into triples and a knowledge graph is constructed, breaking through the limitations of traditional flat label systems and realizing the upgrade of customer portrait from "attribute list" to "semantic network", which can reveal the deep associations between customers and products, services, etc., such as identifying the implicit rule that "customers who buy air purifiers generally pay attention to maternal and child health". Third, the knowledge graph has good scalability and dynamic updating capability, and new customers or interaction events can be added as new nodes or edges in real time, supporting the continuous evolution of customer portrait. Finally, this structured knowledge system provides strong reasoning support for subsequent personalized recommendation, cross-selling, and loss warning, significantly improving the depth of customer understanding and the accuracy of marketing decisions.

[0098] In some embodiments of the present application, the customer behavior sequence data is input into the time sequence fusion transformer model, the long-term dependence and short-term fluctuation characteristics are captured by using the self-attention mechanism in the time sequence fusion transformer model, and the behavior mode is identified, specifically:

[0099] The behavior sequence of the customer in the preset time window is extracted from the fusion data, including page browsing, product adding, order payment, customer service consultation, content liking and search keywords, sorted by timestamp and constructed as multivariate time sequence data;

[0100] The time sequence data is input into the time sequence fusion transformer model, the static features and dynamic features are separated by the gating mechanism, and the dependence weight between different time steps is calculated by using the self-attention layer;

[0101] The behavior mode is identified by analyzing the interpretable attention weight.

[0102] By inputting the customer behavior sequence data into the time sequence fusion transformer (TFT) model, high-precision modeling and interpretable analysis of the dynamic evolution of customer behavior are realized. Specifically, first, the complete behavior trajectory of the customer in the preset time window (such as the last 7 days or 30 days) is extracted from the fused multi-source data, covering page browsing, product adding, order payment, customer service consultation, content liking and search keywords and other interactive events. These behaviors are accurately sorted by timestamp and constructed as multivariate time series data, each time step corresponding to a group of behavior feature vectors, forming a “timeline” representation of customer behavior. Subsequently, the time sequence data is input into the time sequence fusion transformer model, which effectively separates the long-term stable static covariates (such as customer basic attributes) and short-term changing dynamic behavior features by using the built-in gating mechanism (such as GRU or LSTM gating unit), ensuring that the model can utilize customer inherent attributes as context reference and focus on the time sequence evolution law of behavior sequence during modeling. On this basis, the model uses the self-attention mechanism (Self-Attention) to calculate the dependence weight between different time steps, automatically identifies the key behavior nodes and their mutual influence, such as “a user browses high-end mobile phones for 3 times in a row and completes adding on the 5th day”, thereby capturing the complex correlation between long-term interest accumulation and short-term decision triggering. Further, by visualizing and analyzing the weight output by the self-attention layer, the key mode driving customer behavior can be clearly identified, such as “search keywords + adding + consulting customer service” forming a high-conversion path, or “behavior frequency drops sharply after multiple complaints” indicating a risk of loss.

[0103] The time series fusion transformer model overcomes the limitations of traditional time series models (such as ARIMA and simple RNN) in modeling long and short-term dependencies, can process long-span behavior data in parallel, accurately capture the gradual evolution and sudden fluctuations of customer interest, and significantly improve the accuracy of behavior prediction. Second, through the gating mechanism, static and dynamic features are modeled together, enhancing the model's sensitivity to individual differences and making the prediction results more personalized. Third, the self-attention mechanism not only improves the model performance, but also has the natural advantage of interpretability - by analyzing the attention weight distribution, business personnel can intuitively understand the model's judgment basis, such as identifying that "a customer recently frequently searches for Bluetooth earphones" is the key reason for predicting him as a high conversion potential customer, thereby enhancing the model's credibility and feasibility in actual business. Finally, this method supports automatic pattern mining of massive customer behavior paths, can discover group behavior rules (such as typical conversion paths during promotions), and provides data-driven decision support for product optimization, marketing strategy design, and user experience improvement.

[0104] In some embodiments of the present application, a dynamic feature vector containing time evolution information is output to represent the current interest state and future behavior trend of the customer, specifically:

[0105] The time series fusion transformer model outputs a high-dimensional hidden state vector sequence, with each time step corresponding to a dynamic feature vector representing the customer's interest intensity and behavior tendency at that moment;

[0106] The hidden state vector at the final time step is nonlinearly mapped to generate a current interest state vector, which describes the customer's focus at the current time;

[0107] The time series fusion transformer model outputs a probability distribution of multiple future time steps using the built-in prediction head, generating a future behavior trend prediction result.

[0108] The dynamic feature vector output by the deep structure of the Temporal Fusion Transformer (TFT) model contains time evolution information, realizing fine representation of customer behavior state and accurate prediction of future trends. Specifically, during the forward propagation of the model, the TFT generates a high-dimensional hidden state vector sequence through its encoder-decoder architecture, where each hidden state vector at a time step encodes the customer's comprehensive response at that moment in multiple behaviors (such as browsing, adding to cart, searching, etc.), effectively representing the customer's interest intensity, behavior tendency, and association with historical context, forming a continuous evolving "behavior trajectory". Subsequently, for the hidden state vector at the final time step, a nonlinear mapping network (such as a multilayer perceptron) is used to convert it into a "current interest state vector", which describes the customer's focus at the current moment in the form of a high-dimensional semantic vector, such as interest index in a certain category of goods, price sensitivity change, or service consultation tendency, providing accurate state input for real-time personalized recommendation and immediate intervention. At the same time, the prediction head of the TFT model directly outputs the behavior probability distribution at multiple future time steps using historical sequences and context information, such as the probability of completing a purchase within the next 3 days, the next visit time window, and the churn risk score, generating structured future behavior trend prediction results, realizing the transition from "describing the past" to "anticipating the future".

[0109] By outputting the time-sequenced hidden state vector sequence, the model not only captures the instantaneous state of customer behavior but also preserves its dynamic evolution path, providing a data foundation for analyzing the customer interest migration process and improving the depth of customer understanding. Secondly, mapping the final hidden state to a "current interest state vector" realizes the quantitative expression of the customer's real-time intention, enabling the system to respond sensitively to the customer's latest behavior and support millisecond-level personalized services such as real-time pop-up recommendations or intelligent customer service script adjustments. Thirdly, using the TFT model's original multi-step prediction capability to generate future behavior trends avoids the complexity of building additional prediction models in traditional methods, improving system integration efficiency, and its attention mechanism-based prediction has stronger context awareness, significantly outperforming traditional statistical or machine learning models in prediction accuracy. Finally, the dynamic feature vector integrates both "current state" and "future trend" information, forming the core of the "evolution layer" of the customer digital twin, enabling the twin to have time perception and trend deduction capabilities, truly realizing the transition from static portrait to dynamic life.

[0110] In some embodiments of the present application, the customer basic attribute label is weighted and fused with the dynamic feature vector to construct the customer digital twin, specifically:

[0111] The customer basic attribute label is encoded into a fixed-dimensional static feature vector, and a one-hot encoding or embedding representation is used to map it to the same high-dimensional semantic space as the dynamic feature vector;

[0112] An active degree index of the customer is obtained, and the active degree index is calculated based on the frequency of behavior events in the last 30 days, and a formula is A = log(1 + N events ), wherein A represents an active degree score, N events is the total number of valid behaviors of the customer in the statistical period;

[0113] A dynamic fusion weight is calculated according to the active degree score A, wherein a dynamic feature weight calculation formula is w d = σ(k • (A - A0)), wherein σ is a Sigmoid function, k is a gain coefficient, A0 is a baseline active degree threshold, and a static feature weight calculation formula is w s = 1 - w d ;

[0114] The static feature vector v s and the dynamic feature vector v d are weighted and fused to generate a customer comprehensive feature vector v twin , wherein v twin = w s • v s + w d • v d .

[0115] By adaptively weighting and fusing the customer basic attribute label and the dynamic behavior feature vector, a customer digital twin with stability and sensitivity is constructed, and comprehensive, accurate and dynamic representation of the customer state is realized. Specifically, first, the customer basic attribute label (such as age, gender, consumption level, product preference, etc.) extracted from the knowledge graph is converted into a fixed-dimensional static feature vector through one-hot encoding (One-Hot Encoding) or embedding representation (Embedding), and is mapped to the same high-dimensional semantic space as the dynamic feature vector, ensuring that the two types of heterogeneous features can be effectively fused in numerical scale and semantic structure. Subsequently, the active degree index A of the customer is introduced as the basis for adjusting the fusion weight, wherein N events represents the total number of valid behavior events (such as browsing, adding to cart, payment, consultation, etc.) of the customer in the last 30 days, and the logarithmic transformation form can suppress the influence of extremely high-frequency users while reasonably distinguishing between low, medium and high active customers, thereby improving the robustness of the index. Based on the active degree score A, the system dynamically calculates the fusion weight: the dynamic feature weight, wherein σ is a Sigmoid function, k controls the weight change slope, A0 is a preset baseline active degree threshold (such as the median level), and when the active degree of the customer is higher than the threshold, w dApproaching 1, giving recent behavior a higher influence; conversely, w s Dominant, enhancing the stability contribution of static labels. Finally, the customer comprehensive feature vector is generated by linear weighting, which is the core representation of the customer digital twin.

[0116] By fusing static labels and dynamic behavior vectors in a unified semantic space, the problem of "static and dynamic fragmentation" in traditional customer portraits is overcome, making the twin reflect both the long-term attributes of the customer and the short-term behavior changes, improving the completeness and real fit of the portrait. Secondly, the dynamic weighting mechanism based on activity is introduced, realizing the "different for different people" fusion strategy: for high-activity customers (such as frequent shoppers), the system relies more on their latest behavior patterns for judgment, ensuring the timeliness of recommendations and services; for low-activity or silent customers, more historical attributes are relied on for inference, avoiding misjudgment due to sparse data, and enhancing the robustness and personalization ability of the model. Thirdly, the fusion weight is driven by an interpretable mathematical formula, and the parameters k and A0 can be flexibly adjusted according to business needs, supporting the configurability and A / B testing optimization of the strategy, and improving the operability of the system. Finally, the generated customer comprehensive feature vector serves as the unified representation of the digital twin, which can be directly used for downstream tasks such as similar customer clustering, churn prediction, personalized content generation, etc., significantly improving the response accuracy and decision efficiency of the intelligent system.

[0117] In some embodiments of the present application, it also includes:

[0118] The constructed customer digital twin is input into a large language model, and through prompt word engineering, the customer state and trend are converted into natural language description to generate personalized service strategy or risk warning suggestion, and output to CRM system, intelligent customer service platform or marketing automation system through API interface.

[0119] After the customer digital twin is generated, its core state information, including static attribute labels (such as "35-year-old male, first-tier city, high consumption ability"), dynamic behavior characteristics (such as "frequently browsing high-end smart watches in the past week"), and future behavior trend predictions (such as "the probability of purchasing a wearable device within the next 7 days is 82%"), is structured and organized into natural language prompts (Prompts) and input into a pre-trained large language model (such as GPT, Tongyi Qianwen, etc.). Through a carefully designed prompt template, the model is guided to generate natural language descriptions that conform to business logic, such as: "The customer is a high-value user and has recently shown strong interest in smart wearable devices. It is recommended to push new product pre-sale information and provide exclusive coupons," or "The customer has not logged in for two consecutive weeks and has an increasing number of consultation records. The risk of loss is high, and it is recommended that the customer manager actively follow up." The generated personalized service strategies or risk warning suggestions are output in real time to the CRM system, intelligent customer service platform, or marketing automation system through a standardized API interface, driving precise touch and intelligent decision-making.

[0120] Through the semantic generation capability of the large language model, high-dimensional and abstract vector representations are converted into human-readable and business-operable natural language suggestions, greatly improving the explainability and landing efficiency of customer insights, reducing the data analysis threshold, and enabling non-technical personnel to quickly understand customer status and take action. Secondly, the prompt word engineering gives the system high flexibility and controllability, and can guide the model to generate strategy suggestions of different styles and different granularities by adjusting the prompt template, adapting to various application scenarios such as customer service scripts, marketing copy, and management reports. Thirdly, this method realizes the seamless integration of customer digital twin and intelligent decision-making system, forming a closed-loop process of "perception-understanding-decision-making-execution", significantly improving the personalization, automation, and intelligence level of enterprise services. Finally, through the API interface, it realizes the loose coupling of existing business systems, and can be quickly deployed without reconfiguring the original IT architecture, with good scalability and engineering feasibility.

[0121] The third aspect embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to realize the customer twin construction method based on multi-source data fusion and large model driving in any embodiment of the first aspect.

[0122] Figure 2 An example of an entity structure diagram of an electronic device is shown in Figure 2As shown, the electronic device can include a processor 810, a communications interface 820, a memory 830, and a communications bus 840, wherein the processor 810, the communications interface 820, and the memory 830 complete mutual communication through the communications bus 840. The processor 810 can invoke a logic instruction in the memory 830 to execute the customer twin construction method based on multi-source data fusion and large model driving in any embodiment of the first aspect, which includes:

[0123] Step 100, through the RESTful API interface, respectively connect to the enterprise internal system, external platform and Internet of Things device, and collect structured data, unstructured data and semi-structured data in real time.

[0124] Step 200, deploy a feature extraction model locally in each data source, perform distributed training using a federated learning framework, and only upload encrypted local model parameters to a central node. The central node performs aggregation operations on multiple local model parameters to generate a global fusion model.

[0125] Step 300, extract customer basic attribute labels from the cleaned fusion data, including demographic information, consumption ability level, and product preference category, construct a customer knowledge graph based on an entity-relation-entity triple structure, and store the "customer-product" and "customer-service" interaction relationships.

[0126] Step 400, input customer behavior sequence data into a time series fusion transformer model, use the self-attention mechanism in the time series fusion transformer model to capture long-term dependencies and short-term fluctuation characteristics, identify behavior patterns, and output dynamic feature vectors containing time evolution information, which are used to represent the current interest state and future behavior trend of the customer.

[0127] Step 500, weight and fuse the customer basic attribute labels and dynamic feature vectors to construct a customer digital twin.

[0128] Moreover, the logic instructions in the memory 830 described above can be implemented in the form of software function units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the part of the prior art or the part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0129] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the customer twin construction method based on multi-source data fusion and large model driving provided by the above-mentioned method, the method comprising:

[0130] Step 100, through the RESTful API interface, the enterprise internal system, the external platform and the Internet of Things device are connected respectively, and structured data, unstructured data and semi-structured data are collected in real time.

[0131] Step 200, a feature extraction model is deployed locally in each data source, distributed training is performed using a federated learning framework, only local model parameters are encrypted and uploaded to a central node, the central node performs aggregation operation on multiple local model parameters to generate a global fusion model.

[0132] Step 300, customer basic attribute labels are extracted from the cleaned fusion data, the customer basic attribute labels include demographic information, consumption ability level and product preference category, a customer knowledge graph is constructed based on an entity-relation-entity triple structure, and "customer-product" and "customer-service" interaction relationships are stored.

[0133] Step 400, customer behavior sequence data is input into a time series fusion transformer model, a self-attention mechanism in the time series fusion transformer model is used to capture long-term dependence and short-term fluctuation characteristics, identify behavior patterns, and output dynamic feature vectors containing time evolution information, which are used to represent current interest state and future behavior trend of customers.

[0134] Step 500, the customer basic attribute labels and the dynamic feature vectors are weighted and fused to construct a customer digital twin.

[0135] In another aspect, the present application also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the method for constructing a customer twin based on multi-source data fusion and large model driving provided above, which comprises:

[0136] Step 100, connecting to enterprise internal systems, external platforms and Internet of Things devices through RESTful API interfaces respectively, and collecting structured data, unstructured data and semi-structured data in real time.

[0137] Step 200, deploying feature extraction models locally at each data source, performing distributed training using a federated learning framework, and only uploading encrypted local model parameters to a central node, wherein the central node performs aggregation operations on multiple local model parameters to generate a global fusion model.

[0138] Step 300, extracting customer basic attribute labels from the cleaned fusion data, wherein the customer basic attribute labels include demographic information, consumption ability level and product preference category, constructing a customer knowledge graph based on an entity-relation-entity triple structure, and storing "customer-product" and "customer-service" interaction relationships.

[0139] Step 400, inputting customer behavior sequence data into a time series fusion transformer model, capturing long-term dependencies and short-term fluctuations using a self-attention mechanism in the time series fusion transformer model, identifying behavior patterns, and outputting dynamic feature vectors containing time evolution information to represent current interest states and future behavior trends of customers.

[0140] Step 500, weighting and fusing the customer basic attribute labels and dynamic feature vectors to construct a customer digital twin.

[0141] Finally, the present application also provides a non-volatile computer storage medium having stored thereon computer executable instructions which, when executed by a processor, implement the method for constructing a customer twin based on multi-source data fusion and large model driving provided above, which comprises:

[0142] Step 100, connecting to enterprise internal systems, external platforms and Internet of Things devices through RESTful API interfaces respectively, and collecting structured data, unstructured data and semi-structured data in real time.

[0143] Step 200, deploying feature extraction models locally at each data source, performing distributed training using a federated learning framework, and only uploading encrypted local model parameters to a central node, wherein the central node performs aggregation operations on multiple local model parameters to generate a global fusion model.

[0144] Step 300, extracting customer basic attribute labels from the cleaned fusion data, the customer basic attribute labels including demographic information, consumption ability level and product preference category, constructing a customer knowledge graph based on the entity-relation-entity triple structure, and storing the "customer-product" and "customer-service" interaction relationships.

[0145] Step 400, inputting the customer behavior sequence data into the time sequence fusion transformer model, capturing long-term dependence and short-term fluctuation characteristics by using the self-attention mechanism in the time sequence fusion transformer model, identifying behavior patterns, and outputting dynamic feature vectors containing time evolution information for representing the current interest state and future behavior trend of the customer.

[0146] Step 500, weighting and fusing the customer basic attribute labels and the dynamic feature vectors to construct a customer digital twin.

[0147] The places not mentioned in the present application can be realized by using or referring to the existing technology.

[0148] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment mainly describes the difference from other embodiments.

[0149] The above only describes the embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various changes and modifications to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A customer twin construction method based on multi-source data fusion and large model driving, characterized in that, The application relates to a customer digital twin construction method based on federated learning, comprising the following steps: connecting enterprise internal systems, external platforms and Internet of Things devices through RESTful API interfaces to collect structured data, unstructured data and semi-structured data in real time; locally deploying feature extraction models on each data source, performing distributed training using a federated learning framework, and uploading only local model parameters to a central node after encryption, wherein the central node performs an aggregation operation on multiple local model parameters to generate a global fusion model; extracting customer basic attribute labels from the cleaned fusion data, wherein the customer basic attribute labels include demographic information, consumption capacity level and product preference category, constructing a customer knowledge graph based on an entity-relation-entity triple structure, and storing "customer-product" and "customer-service" interaction relationships; inputting customer behavior sequence data into a time series fusion transformer model, capturing long-term dependence and short-term fluctuation characteristics by using a self-attention mechanism in the time series fusion transformer model, identifying behavior patterns, and outputting a dynamic feature vector containing time evolution information to represent the current interest state and future behavior trend of the customer; weighting and fusing the customer basic attribute labels and the dynamic feature vector to construct a customer digital twin.

2. The method of claim 1, wherein the method is characterized by, The structured data includes customer basic information and order records, the unstructured data includes user comments and customer service audio-to-text content, and the semi-structured data includes user behavior logs and click stream data.

3. The method of claim 1, wherein the method is characterized by, The local feature extraction model is initialized on each data source side, and the local feature extraction model includes a multilayer perceptron subnetwork for processing structured data, a lightweight BERT embedding layer for processing unstructured text data, and a projection layer for aligning output dimensions; each data source uses local data to locally train the local feature extraction model, updates the model parameters, and does not upload original data during the training process; the updated local model parameters are uploaded to the central node after encryption and a federated averaging algorithm or a momentum-based aggregation algorithm is executed to generate an updated global fusion model; the global fusion model parameters are distributed to each local node for the next round of training iteration until the model converges; based on the converged global fusion model, cross-domain feature extraction and semantic alignment are performed on new samples. The customer basic attribute labels are extracted from the cleaned fusion data, and the specific steps are as follows:

4. The method of claim 1, wherein the method is characterized by, customer static attribute fields are extracted from the fusion data, including age, gender, region, registration duration, annual total consumption, order frequency and membership level; the static attribute fields are classified and labeled based on a preset rule or a clustering algorithm to generate basic labels: based on the basic labels, customers and their associated entities are organized into triples in the form of "subject-predicate-object"; the triples are imported into a graph database to construct a customer knowledge graph. ​ 5. The method of claim 1, wherein the method is characterized by, The customer behavior sequence data is input into the time sequence fusion transformer model, long-term dependence and short-term fluctuation characteristics are captured by using a self-attention mechanism in the time sequence fusion transformer model, and a behavior pattern is identified, specifically: The behavior sequence of the customer in a preset time window is extracted from the fusion data, including page browsing, product adding, order payment, customer service consultation, content liking, and search keywords, sorted by timestamp and constructed as multivariate time sequence data; The time sequence data is input into the time sequence fusion transformer model, the static features and dynamic features are separated by a gating mechanism, and the dependence weight between different time steps is calculated by using a self-attention layer; The behavior pattern is identified by analyzing the interpretable attention weight.

6. The method of claim 5, wherein the method is characterized by, The output includes a dynamic feature vector containing time evolution information, which is used to represent the current interest state and future behavior trend of the customer, specifically: The time sequence fusion transformer model outputs a high-dimensional hidden state vector sequence, each time step corresponds to a dynamic feature vector, representing the interest intensity and behavior tendency of the customer at that moment; The hidden state vector at the final time step is nonlinearly mapped to generate a current interest state vector, which is used to describe the focus of the customer at the current moment; The prediction head of the time sequence fusion transformer model is used to output the probability distribution of multiple future time steps, and the future behavior trend prediction result is generated.

7. The method of claim 1, wherein the method is characterized by, The customer basic attribute label and the dynamic feature vector are weighted and fused to construct a customer digital twin, specifically: The customer basic attribute label is encoded into a fixed-dimensional static feature vector, using one-hot encoding or embedding representation, and mapped to the same high-dimensional semantic space as the dynamic feature vector; Obtaining a customer activity index, the customer activity index being calculated based on a frequency of behavior events in the last 30 days, a formula being A=log(1+N events ), wherein A represents an activity score, N events is a total number of valid behaviors of the customer in a statistical period; The fusion weight is dynamically calculated according to the activity score A, wherein a dynamic feature weight calculation formula is w d =σ(k•(A-A0)), wherein σ is a Sigmoid function, k is a gain coefficient, A0 is a reference activity threshold, and a static feature weight calculation formula is w s =1-w d ; The static feature vector v s and the dynamic feature vector v d are weightedly fused to generate a customer comprehensive feature vector v twin , wherein v twin = w s • v s + w d • v d .

8. The method of claim 1 to 7, wherein, Further comprising: The constructed customer digital twin is input into a large language model, the customer state and trend are converted into natural language description through prompt word engineering, personalized service strategy or risk warning suggestion is generated, and is output to the CRM system, intelligent customer service platform or marketing automation system through API interface.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the customer twin construction method based on multi-source data fusion and large model driving according to any one of claims 1-8.

10. A non-transitory computer storage medium having stored thereon computer- executable instructions which, when executed by a computer, cause the computer to perform: The computer executable instructions are executed by the processor to realize the customer twin construction method based on multi-source data fusion and large model driving according to any one of claims 1-8.

Images