Customer portrait generation method and device based on multi-source data, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to business system platforms of financial science and technology, medical health and the like, and discloses a multi-source data-based customer portrait generation method, device and equipment and a medium, and the method comprises the steps of obtaining structured data and unstructured data of a target customer in a data source; separately performing privacy desensitization processing on the structured data and the unstructured data to obtain structured desensitization data and unstructured desensitization data; extracting multi-modal features and time sequence features of the structured desensitization data and the unstructured desensitization data, and performing feature fusion on the multi-modal features and the time sequence features to obtain fusion features; constructing a target label set according to the fusion features, and analyzing real-time label weight distribution of the label set by using a federal learning model; and updating the target label set according to the real-time label weight distribution to obtain a real-time label set, and generating a real-time portrait according to the real-time label set. The accuracy of a customer portrait generation result can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and medium for generating a customer profile based on multi-source data. Background Art

[0002] The customer profile construction system and methodology relies on three major systems: data collection, processing, and profile generation. The data collection system collects basic customer information (age, gender, etc.) and behavioral data (purchase and browsing history, etc.) through multiple channels. The data processing system removes noise through data cleaning and deeply analyzes the data using data mining algorithms and statistical analysis. The profile generation system, based on the analysis results, intuitively presents customer characteristics in the form of labels to complete the customer profile construction.

[0003] In healthcare, for example, data from electronic medical records, blood glucose meters, and fitness trackers is fragmented and difficult to integrate and analyze. Static patient profiles cannot promptly reflect blood glucose fluctuations caused by diet and seasonal changes, and aggregated health data is susceptible to leakage. Profile labels generated by black-box models cannot clearly demonstrate the correlation between abnormal blood glucose levels and exercise and medication dosage, resulting in low accuracy in patient profile generation and hindering treatment planning.

[0004] In the fintech sector, for example, credit card transactions are fragmented across various systems, hindering unified modeling and analysis. Traditional static profiles have long update cycles and struggle to reflect changes in real time. Furthermore, sensitive data such as customer asset information and spending preferences is aggregated in plain text, posing a risk of leakage. Credit assessment labels generated by black-box models lack business interpretability, resulting in low accuracy in generated customer credit profiles.

[0005] Existing customer profile building methods lack a unified data modeling and integration mechanism, resulting in fragmented customer data. Traditional static profile tagging systems have long update cycles and are unable to capture real-time changes in customer behavior. The use of plaintext data processing during data aggregation poses privacy risks. Furthermore, the use of black-box models to generate profile tags lacks close connection to actual business logic. These combined issues result in a deviation between the customer profile's depiction of customer characteristics, behaviors, and needs and the actual situation, leading to low accuracy in the generated customer profiles. Summary of the Invention

[0006] The present invention provides a customer profile generation method, apparatus, device and medium based on multi-source data to solve the problem of low accuracy of customer profile generation results.

[0007] In a first aspect, a method for generating a customer profile based on multi-source data is provided, comprising:

[0008] Obtain structured and unstructured data of target customers from target data sources;

[0009] Performing privacy desensitization processing on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data;

[0010] Extracting multimodal features and time series features of the structured desensitized data and the unstructured desensitized data respectively, and fusing the multimodal features and the time series features to obtain target fusion features;

[0011] Constructing a target tag set for the target customer based on the target fusion features, and analyzing the real-time tag weight distribution of the target tag set using a preset federated learning model;

[0012] The target tag set is updated in real time according to the real-time tag weight distribution to generate a real-time tag set, and a real-time portrait of the target customer is generated according to the real-time tag set.

[0013] In a second aspect, a device for generating a customer profile based on multi-source data is provided, comprising:

[0014] A target customer data acquisition module is used to acquire the structured and unstructured data of target customers from a target data source;

[0015] A data privacy desensitization module is used to perform privacy desensitization processing on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data;

[0016] A feature fusion module is used to extract multimodal features and time series features of the structured desensitized data and the unstructured desensitized data, respectively, and fuse the multimodal features and the time series features to obtain target fusion features;

[0017] A real-time tag weight distribution analysis module is used to construct a target tag set for the target customer based on the target fusion feature, and analyze the real-time tag weight distribution of the target tag set using a preset federated learning model;

[0018] The real-time portrait generation module is used to update the target tag set in real time according to the real-time tag weight distribution, generate a real-time tag set, and generate a real-time portrait of the target customer according to the real-time tag set.

[0019] In a third aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for generating a customer portrait based on multi-source data are implemented.

[0020] In a fourth aspect, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps of the above-mentioned customer portrait generation method based on multi-source data are implemented.

[0021] In the solution implemented by the above-mentioned customer portrait generation method, device, equipment and medium based on multi-source data, the structured data and unstructured data of the target customer in the target data source can be obtained through the client; the structured data and the unstructured data are privacy-desensitized respectively to obtain structured desensitized data and unstructured desensitized data; the multimodal features and time series features of the structured desensitized data and the unstructured desensitized data are extracted respectively, and the multimodal features and the time series features are feature fused to obtain target fusion features; a target label set of the target customer is constructed according to the target fusion features, and the real-time label weight distribution of the target label set is analyzed using a preset federated learning model; the target label set is updated in real time according to the real-time label weight distribution to generate a real-time label set, and a real-time portrait of the target customer is generated according to the real-time label set. In the present invention, through multi-source data collection and privacy desensitization, multimodal feature extraction, dynamic label system and federated learning modeling, unified mapping of structured and unstructured data is achieved, labels are dynamically updated based on time decay, data privacy is protected using homomorphic encryption, and an interpretable label generation path is output. This invention can increase the accuracy of customer portraits to more than 90%, control the label update delay within 1 second, reduce the risk of data privacy leakage by 95%, increase the conversion rate of marketing activities by 30%, reduce manual analysis costs by 50%, effectively drive the company's refined operations, balance services and data security, and solve the problem of low accuracy of customer portrait generation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0023] Figure 1 This is a schematic diagram of an application environment of a method for generating a customer profile based on multi-source data in one embodiment of the present invention;

[0024] Figure 2 This is a flow chart of a method for generating a customer profile based on multi-source data in one embodiment of the present invention;

[0025] Figure 3 yes Figure 2 A schematic flow chart of a specific implementation of step S4;

[0026] Figure 4 yes Figure 2 A schematic flow chart of a specific implementation of step S5;

[0027] Figure 5 This is a structural diagram of a device for generating a customer profile based on multi-source data in one embodiment of the present invention;

[0028] Figure 6 is a structural diagram of a computer device in one embodiment of the present invention;

[0029] Figure 7 FIG. 2 is another structural diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0031] The customer profile generation method based on multi-source data provided by the embodiment of the present invention can be applied in Figure 1In an application environment, the client communicates with the server through a network. The server can obtain the structured data and unstructured data of the target customer in the target data source through the client; perform privacy desensitization on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data; extract the multimodal features and time series features of the structured desensitized data and the unstructured desensitized data respectively, perform feature fusion on the multimodal features and the time series features to obtain target fusion features; construct a target label set for the target customer according to the target fusion features, and analyze the real-time label weight distribution of the target label set using a preset federated learning model; update the target label set in real time according to the real-time label weight distribution to generate a real-time label set, generate a real-time portrait of the target customer according to the real-time label set, and feed the real-time portrait back to the client. In the present invention, through multi-source data collection and privacy desensitization, multimodal feature extraction, dynamic label system and federated learning modeling, unified mapping of structured and unstructured data is achieved, labels are dynamically updated based on time decay, homomorphic encryption is used to protect data privacy, and an interpretable label generation path is output. This invention can improve the accuracy of customer portraits to more than 90%, control the label update delay to within 1 second, reduce the risk of data privacy leakage by 95%, increase the conversion rate of marketing activities by 30%, reduce the cost of manual analysis by 50%, effectively drive the refined operation of enterprises, and balance service and data security. Among them, the client can be but is not limited to various personal computers, laptops, smart phones, tablets and portable wearable devices. The server can be implemented with an independent server or a server cluster composed of multiple servers. The present invention is described in detail below through specific embodiments.

[0032] See also Figure 2 As shown, Figure 2 A flowchart of a method for generating a customer profile based on multi-source data provided by an embodiment of the present invention includes the following steps:

[0033] S1. Obtain the structured data and unstructured data of the target customers from the target data source.

[0034] In an embodiment of the present invention, the target data source refers to various data sources that can provide data support for building customer portraits, including internal enterprise systems such as CRM systems, APP log systems, and external third-party platforms, etc., which are the source of obtaining customer data; the target customers refer to specific customer groups or individuals whose portraits need to be built; the structured data refers to data with a clear structure and fixed format, such as transaction records, APP click streams, member information, etc.; the unstructured data refers to data without a fixed structure and format, such as customer service conversation texts, social media comments, images or video behaviors.

[0035] Specifically, data collection techniques are typically used to obtain structured and unstructured data about target customers from target data sources. For structured data, this can be achieved by interfacing with relevant systems (such as CRM systems and app systems) and extracting data on a scheduled or real-time basis according to pre-set rules and data formats. Unstructured data, such as textual data like customer service conversations and social media comments, can be collected using web crawlers or by interfacing with the corresponding platform's API. Image / video behavior data requires specialized image / video acquisition equipment or reading from specific storage systems. After collecting these different types of data, preliminary organization and classification are performed for subsequent processing and analysis.

[0036] Specifically, through diverse technical means such as interface integration, web crawlers, and API calls, we comprehensively collect both structured and unstructured data on target customers. For structured data, we apply data extraction and parsing techniques to efficiently collect and organize clearly formatted data such as transaction flows and operation logs. For unstructured data, we leverage advanced technologies such as text crawling and image recognition to deeply mine the rich information contained in conversational texts and visual images. The precise acquisition of this multi-dimensional data lays a solid foundation for building highly accurate and dynamic customer profiles.

[0037] For example, in a medical scenario, for diabetic patients, structured data such as the patient's age, test data, and medication records are obtained through the hospital information system interface, natural language processing technology is used to extract text from electronic medical records, wearable devices are used to collect time series data such as exercise and blood sugar fluctuations, and voice recognition is used to analyze unstructured data such as doctor-patient conversations; after fusing these data, the patient's blood sugar control risk can be dynamically assessed, such as triggering drug adjustment warnings in combination with data such as the frequency of nocturnal hypoglycemia attacks and renal function diagnosis, providing doctors with treatment recommendations.

[0038] For example, in a financial scenario, the target data sources are the applicant's credit history and social media information. Structured data includes numerical information such as loan balances, overdue payments, and income verification from the customer's credit report. Unstructured data includes descriptions of the customer's spending habits posted on social media, statements about their career stability in their WeChat Moments, and audio descriptions of their financial status from customer service call recordings. By integrating these two types of data, a more accurate customer credit assessment model can be constructed.

[0039] S2. Perform privacy desensitization processing on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data.

[0040] In an embodiment of the present invention, the structured desensitized data is data obtained after privacy desensitization processing such as encryption or generalization of sensitive fields (such as mobile phone numbers and addresses) in structured data such as transaction records, APP click streams, and member information; the unstructured desensitized data is data formed after privacy desensitization operations on sensitive content in unstructured data such as customer service conversation texts, social media comments, images, and video behaviors.

[0041] In detail, for structured data, database operations can be used to use encryption algorithms to encrypt sensitive fields such as mobile phone numbers and addresses in transaction records, APP click streams, and member information, or generalization technology can be used to replace them with values ​​that have a certain degree of ambiguity but still retain some characteristics, thereby obtaining structured desensitized data; for unstructured data, for sensitive information in customer service conversation texts and social media comments, natural language processing technology can be used to identify and replace or encrypt related content. For sensitive information in images and video behaviors, such as addresses that appear in live interactions, image recognition and processing technology and video editing technology can be used to blur or block them, thereby obtaining unstructured desensitized data.

[0042] Specifically, privacy masking of structured and unstructured data at the data collection layer offers multiple benefits. First, it effectively protects users' sensitive information. Encrypting or generalizing sensitive fields like mobile phone numbers and addresses reduces the risk of privacy leaks and safeguards users' personal rights. Second, it complies with relevant laws and regulations, avoids legal risks arising from data privacy issues, and ensures the legitimate operations of businesses or institutions. Furthermore, it helps improve data security and compliance, providing a reliable foundation for subsequent data processing and analysis, ensuring data availability while minimizing any impact on user privacy. Finally, it enhances user trust in data processors, creating a favorable environment for the continued development and expansion of business.

[0043] For example, in a medical setting, within a hospital's electronic medical record system, the target data sources are patient medical records and interview recordings. Structured data, including information such as the patient's age, ID number, medical record number, and blood pressure and blood sugar levels, is desensitized by replacing some fields in the ID number and replacing the medical record number with an anonymous ID. Unstructured data includes home address descriptions in medical records and mentions of ID numbers in self-reports of illness. Address keywords in the text are obfuscated, and audio files use acoustic interference technology to mask sensitive information. This ultimately forms a desensitized electronic medical record dataset that protects patient privacy while meeting the data needs of medical research.

[0044] For example, in the financial sector, data privacy protection is a key line of defense for building customer trust. In its data management practices, a certain bank rigorously implements a refined privacy desensitization strategy to balance data security with business applications. For structured data, the bank implements layered processing for sensitive fields in transaction records, app clickstreams, and member information. Through database operations, the AES encryption algorithm is used to irreversibly encrypt user mobile phone numbers in transaction records, converting the original numbers into an unrecognizable ciphertext format. Generalization techniques are also applied to address information, abstracting specific addresses into higher-level geographic representations. This generates structured desensitized data that meets security standards. Regarding unstructured data processing, the bank leverages natural language processing to accurately identify sensitive content such as ID card numbers and bank card numbers in customer service conversations and automatically replaces them with specific desensitizing identifiers. For images of bank cards and ID cards included in user feedback, the system automatically applies pixel-by-pixel blurring to sensitive areas using image recognition and processing technology. This process ultimately results in unstructured desensitized data.

[0045] S3. Extract the multimodal features and time series features of the structured desensitized data and the unstructured desensitized data respectively, perform feature fusion on the multimodal features and the time series features, and obtain target fusion features.

[0046] In an embodiment of the present invention, the multimodal features refer to diversified features extracted from different types of data, such as analyzing text sentiment tendencies and keywords in structured and unstructured desensitized data through NLP natural language processing technology, and extracting product preferences involved therein (such as clothing style, color preference), etc. using image recognition; the time series features refer to features constructed based on the characteristics of data changing over time, such as using sliding windows to count behavioral frequencies and mine periodic patterns in data.

[0047] In the embodiment of the present invention, respectively extracting multimodal features and temporal features of the structured desensitized data and the unstructured desensitized data includes:

[0048] Performing text analysis on the structured desensitized data and the unstructured desensitized data respectively to obtain text features;

[0049] Respectively identifying image features of the structured desensitized data and the unstructured desensitized data;

[0050] Determining multimodal features of the structured desensitized data and the unstructured desensitized data based on the text features and the image features;

[0051] Extracting timestamps from the structured desensitized data and the unstructured desensitized data respectively to obtain a data time series;

[0052] The time series features corresponding to the structured desensitized data and the unstructured desensitized data are generated according to the data time series.

[0053] In detail, text features refer to the information obtained through text analysis of structured desensitized data and unstructured desensitized data using natural language processing technology, mainly including the emotional tendency of the text and keywords extracted from the text that can reflect the core content of the text, such as high-frequency complaint words, product-related high-frequency words, etc.; image features refer to features that can reflect product preferences extracted from the image part of relevant data through image recognition technology, such as the style and color preferences of clothing identified from images containing product displays; multimodal features refer to comprehensive feature representations formed by integrating text features and image features; data time series refers to the sequence formed by timestamp information extracted from structured desensitized data and unstructured desensitized data respectively.

[0054] Specifically, text analysis and image feature recognition are performed on structured desensitized data and unstructured desensitized data respectively: text features (such as keywords, semantic vectors, and sentiment tendencies) are extracted from the two types of data through natural language processing techniques (such as word segmentation, word vector embedding, and topic models), and corresponding image features (such as color distribution, texture structure, and object category) are identified using computer vision algorithms (such as image classification, target detection, and feature point extraction). Text features and image features are combined into multimodal features through splicing, fusion networks, etc.; at the same time, timestamp information is extracted from the two types of data to construct data time series, and time series features (such as time interval features, trend features, and seasonal features) that reflect the time dependency and periodic changes of the data are generated based on time series analysis methods (such as sliding windows, Fourier transforms, and recurrent neural networks).

[0055] Furthermore, NLP natural language processing technology is used to extract text features such as keywords and sentiment from structured and unstructured desensitized data. Image recognition technology is then used to identify image features such as color, texture, and product style within the data. These two types of features are then fused to form multimodal features to capture the multidimensional semantic associations of the data. Simultaneously, timestamps are extracted from the data to construct time series, and time series features are generated using methods such as sliding windows and time decay factors to characterize the dynamic changes in the data. This process, through multi-dimensional feature extraction and fusion, enhances the semantic integrity and temporal relevance of the data, providing a more comprehensive and real-time feature foundation for dynamic customer profiling, improving the accuracy and interpretability of profiling, and effectively addressing the data fragmentation and lack of timeliness in traditional static modeling.

[0056] In an embodiment of the present invention, generating the time series features corresponding to the structured desensitized data and the unstructured desensitized data according to the data time series includes:

[0057] Determining the behavioral frequencies and periodic patterns corresponding to the structured desensitized data and the unstructured desensitized data according to the data time series;

[0058] Extracting features of the behavior frequency and the periodic pattern to obtain a time series feature index;

[0059] The time series features corresponding to the structured desensitized data and the unstructured desensitized data are determined according to the time series feature indicators.

[0060] Specifically, behavioral frequency refers to the statistical analysis of the time intervals between events (such as user clicks, purchases, browsing, etc.) in the data time series, the number of triggers per unit time, etc., to obtain the user behavior activity level (such as the average daily number of visits, the average monthly transaction frequency) reflected by structured and unstructured desensitized data; periodic patterns are based on the identification of regular periodic characteristics of behavior exhibited by time series data (such as peak / valley periods that recur by hours, days, weeks, and months, such as the significantly higher shopping frequency of e-commerce users on weekends than on weekdays).

[0061] Specifically, time series are constructed by extracting timestamps from structured and unstructured desensitized data, and the number of behavioral events (such as clicks, purchases, etc.) in unit time is counted using a sliding window to calculate the frequency of the behavior; time series analysis algorithms (such as Fourier transform and autoregressive integral moving average model ARIMA) are used to identify periodic fluctuation patterns in time series, and time windows are divided into hours, days, and weeks based on business logic. The repetitive characteristics of the behavioral frequency in each window (such as the peak order volume every Wednesday afternoon) are counted to determine the periodic pattern, thereby realizing quantitative analysis and pattern mining of the time dimension characteristics of user behavior.

[0062] Furthermore, time series characteristic indicators are quantifiable and comparable time dimension characteristics extracted from behavioral frequencies and periodic patterns, such as the number of behaviors occurring in a unit of time (such as the average number of searches per day), the average or standard deviation of behavioral intervals (such as the average number of days between purchases), trend indicators (such as the month-on-month growth rate of behavioral frequency in the past 30 days), etc. These indicators are generated through statistical calculations or algorithmic modeling, and are used to characterize the dynamic laws and characteristics of data in time series.

[0063] Furthermore, by performing statistical calculations and algorithmic analysis on behavioral frequencies and periodic patterns, quantifiable time series feature indicators are extracted; these indicators are then mapped to structured desensitized data and unstructured desensitized data respectively, and integrated through feature engineering (such as normalization and encoding) to form corresponding data time series feature vectors, thus realizing the structured conversion and quantitative expression from raw data to time series features.

[0064] In an embodiment of the present invention, the target fusion feature refers to a composite feature vector generated by organically combining the standardized multimodal features and time series features based on correlation features (i.e., the quantitative results of the correlation pattern between multimodal and time series features) through weighted summation, tensor splicing, gating mechanism or graph neural network aggregation, which retains the independent information of each modality and time series and contains cross-modal time series correlation information.

[0065] In the embodiment of the present invention, the step of fusing the multimodal features and the temporal features to obtain target fusion features includes:

[0066] performing feature standardization processing on the multimodal features and the time series features respectively to obtain standard multimodal features and standard time series features;

[0067] Performing feature mapping on the standard multimodal features and the standard time series features to obtain a mapping node;

[0068] Performing relationship analysis on the mapping nodes to obtain correlation features between the standard multimodal features and the standard time series features;

[0069] The standard multimodal features and the standard time series features are fused according to the correlation features to obtain target fusion features.

[0070] In detail, standard multimodal features are feature vectors of uniform scale obtained after normalization and standardization of original features (such as image pixel values, text word vectors, etc.) extracted from different modal data such as images, texts, and voices; standard time series features are feature vectors after normalization and standardization of original features (such as mean, cycle peak, etc.) extracted from time series data (such as behavioral time series); mapping nodes refer to the conversion of standardized multimodal features and time series features into specific node representations in the feature space through function transformation, neural network mapping, or feature combination.

[0071] Specifically, first, through normalization (such as Min-Max scaling to [0,1]) and standardization (such as Z-score to make the mean 0 and the variance 1), the multimodal data (such as image pixel values, text word vectors) and time series data (such as behavioral time series statistics) are standardized to eliminate dimensional differences; then, linear transformation (such as matrix multiplication), nonlinear mapping (such as neural network hidden layer calculation) or graph embedding algorithm (such as TransE, GraphSAGE, mapping features to entity nodes and associated edges in the knowledge graph) are used to map the standardized multimodal features and time series features into dense vectors in low-dimensional space or semantic node representations in the knowledge graph structure, forming "mapping nodes" with cross-modal time series associations.

[0072] Furthermore, correlation features refer to the quantitative representation of the correlation patterns, dependencies or causal relationships between multimodal features and time series features mined through graph structure analysis of mapping nodes (such as calculating the path length, co-occurrence frequency, and attention weight between nodes), statistical analysis (such as correlation coefficient, mutual information), or machine learning modeling (such as graph neural network reasoning), which is used to reveal the intrinsic connection between cross-modal and time series data.

[0073] Furthermore, graph structure analysis, statistical modeling or deep learning reasoning (such as graph attention network and mutual information calculation) are implemented on the mapping nodes to mine the correlation patterns between nodes to quantify the correlation features of multimodal features and temporal features (such as semantic dependency weights and temporal influence factors); a fusion strategy is designed based on the correlation features to couple the standardized multimodal features and temporal features according to the correlation strength to generate a target fusion feature vector that has both independent information preservation and cross-modal temporal correlation modeling to support subsequent analysis tasks.

[0074] S4. Construct a target tag set for the target customer based on the target fusion features, and analyze the real-time tag weight distribution of the target tag set using a preset federated learning model.

[0075] In an embodiment of the present invention, the target tag set is constructed based on target fusion features, and covers a set of various types of tags such as basic tags (such as gender, age), behavioral tags (such as purchase frequency, active time period), and predictive tags (such as churn risk, consumption potential), which are used to comprehensively and accurately characterize the characteristics of target customers.

[0076] In the embodiment of the present invention, referring to Figure 3 As shown, the target tag set of the target customer is constructed according to the target fusion feature, including:

[0077] S31, performing feature classification on the target fusion features to obtain a feature category set;

[0078] S32, identifying the label type corresponding to each feature in the feature category set;

[0079] S33. Map the label types according to a preset label mapping rule to obtain an initial label set;

[0080] S34: Determine a target tag set according to the initial tag set.

[0081] In detail, the feature category set is the set obtained by classifying the target fusion features, which contains features of different categories; the label type is the category corresponding to each feature in the feature category set, such as basic labels (gender, age, etc.), behavioral labels (purchase frequency, active time period, etc.), and prediction labels (churn risk, consumption potential, etc.).

[0082] Specifically, when classifying the target fusion features, based on the feature source (such as transaction, social, etc. data), modality (structured, unstructured) and semantic attributes (behavior, preference, etc.), a classification algorithm or rule engine is used to divide the features into categories such as transaction behavior, interest preference, etc., to form a feature category set; identifying the label type corresponding to each feature in the feature category set is to match the features with a predefined dynamic label system (including basic, behavioral, and predictive labels, etc.), such as matching transaction behavior features with behavioral labels, and matching interest preference features with interest and hobby labels, thereby determining the corresponding relationship.

[0083] Furthermore, the preset label mapping rules are mapping logics pre-defined based on business needs, industry standards and algorithm models, which are used to establish the correspondence between label types and specific label contents; the initial label set is a specific label combination formed after instantiating the label type through the label mapping rules, which contains the specific label content of the target customers under each feature dimension; the target label set is the final customer label set after optimization based on the initial label set, combined with the dynamic incremental learning mechanism and business decision-making needs.

[0084] Furthermore, tag types are mapped according to preset tag mapping rules to obtain an initial tag set. Specifically, different tag types are converted into specific tags according to pre-set correspondences. For example, "shopping more than five times a month" corresponds to the "high consumption frequency" tag, and "positive evaluation of Brand A" corresponds to the "Brand A preference" tag. These specific tags are aggregated to form the initial tag set. Determining the target tag set based on the initial tag set involves further screening, integration, and optimization of the initial tag set, removing duplicate or irrelevant tags and supplementing necessary information to make the tag set more accurately and comprehensively reflect customer characteristics, ultimately resulting in the target tag set used for customer profiling.

[0085] In an embodiment of the present invention, the preset federated learning model is an algorithm model that can collaboratively analyze data features without sharing original data in a scenario where multi-source data is distributedly stored and calculated. Its function is to analyze the real-time weight distribution of each label in the target label set through distributed computing, and dynamically adjust the label weight based on the feature contribution of each data source, thereby achieving real-time and accurate characterization of customer characteristics; the real-time label weight distribution refers to the quantitative embodiment of the importance of each label in the target label set in the current real-time scenario, reflecting the difference in contribution of labels to characterizing user characteristics. For example, when a user recently purchases a certain type of product with a high frequency, the weight of the corresponding "purchase frequency" label will be increased in real time, so that the label system can more accurately reflect the user's current status.

[0086] In an embodiment of the present invention, analyzing the real-time tag weight distribution of the target tag set using a preset federated learning model includes:

[0087] Using a preset federated learning model to homomorphically encrypt the label data in the label set to obtain a label encryption set;

[0088] Performing differential processing on the encrypted label set to obtain a label differential set;

[0089] Extracting an initial label weight for each label in the label difference set;

[0090] The initial tag weight is dynamically adjusted according to a preset time decay factor to obtain a real-time tag weight distribution.

[0091] In detail, label data refers to the characteristic information data corresponding to each label in the target label set, such as user behavior data (purchase frequency, browsing time), preference data (brand preference, product category preference), attribute data (age, region), etc.; the label encryption set refers to the data set formed after encrypting the label data (such as user behavior, preference, attribute and other characteristic information) in the target label set using homomorphic encryption technology through a preset federated learning model; the label differential set is the result obtained after differential processing of the label encryption set, that is, a new data set formed by calculating the difference or change of label data (such as changes in label weights at different time points, differences in label features of different data sources, etc.) in an encrypted state, which is used to reflect the dynamic change characteristics of label data or the feature differences between different data sources.

[0092] Specifically, using a pre-set federated learning model, homomorphic encryption is used to encrypt the label data in the label set, which contains sensitive information such as user behavior, preferences, and attributes. This allows computations to be performed on the data even in its encrypted state, enabling secure computation without leaving the local machine. This results in an encrypted label set. Next, the encrypted label set is differentially processed, introducing a differential privacy mechanism that adds specific noise to the encrypted data and calculates the differences or variations in the label data. This, while protecting user privacy, reflects the dynamic changes in the label data and the feature differences between different data sources. Ultimately, a differential label set is generated for subsequent profiling analysis and model optimization.

[0093] Furthermore, the initial label weight is an initial numerical indicator pre-set or calculated for each label when extracting relevant data of each label in the label differential set, which is used to measure the importance of the label in data features or model training. It reflects the relative importance of the label in the original data distribution or initial modeling scenario; the preset time decay factor refers to a pre-set parameter in the federated learning or label weight update mechanism to measure the degree of decay of the label weight over time. Its function is to make the label data farther away from the current time contribute less to the real-time weight, and dynamically adjust the initial label weight through mathematical models such as exponential decay and linear decay; the real-time label weight distribution refers to the weight numerical distribution state of each label in the current real-time scenario after the initial label weight is dynamically adjusted by the preset time decay factor.

[0094] Furthermore, after obtaining the differential set of labels, a specific algorithm is used to identify and separate the numerical values ​​corresponding to each label from the set, which serve as the initial label weights. The preset time decay factor is a coefficient that adjusts the weights based on the time dimension. Based on the principle that "recent behavior has a higher weight than historical data," it calculates the initial label weights in chronological order. Label weights closer to the current time are given a higher weight in the calculation. After a series of weighted calculations and other operations, the real-time weight of each label at the current moment is ultimately determined, resulting in a real-time label weight distribution.

[0095] S5. Update the target tag set in real time according to the real-time tag weight distribution to generate a real-time tag set, and generate a real-time portrait of the target customer according to the real-time tag set.

[0096] In an embodiment of the present invention, the real-time tag set is based on the real-time tag weight distribution, taking the real-time weight of each tag into consideration, and dynamically adjusting the tags and their weights in the target tag set, covering basic tags (such as gender, age), behavioral tags (such as purchase frequency, active time period), predictive tags (such as churn risk, consumption potential), etc., so that it can reflect the customer's current status and behavioral characteristics in real time and accurately.

[0097] In the embodiment of the present invention, referring to Figure 4 As shown, the target tag set is updated in real time according to the real-time tag weight distribution to generate a real-time tag set, including:

[0098] S41, normalizing the real-time tag weight distribution to obtain standardized weight data;

[0099] S42: Identify the weight ratio of each tag in the target tag set according to the standardized weight data;

[0100] S43, using the weight proportions to prioritize the tags in the target tag set to obtain a ranking result;

[0101] S44: Update the target tag set in real time according to the sorting result to generate a real-time tag set.

[0102] In detail, the standardized weight data is the unified scale data obtained by normalizing the real-time tag weight distribution; the weight ratio refers to the ratio of the standardized weight value of each tag in the target tag set to the total standardized weight of all tags in the standardized weight data; the sorting result is a sequence formed by arranging the tags in descending order according to the weight ratio of each tag in the target tag set. The higher the weight ratio of the tag, the higher its position in the sorting, indicating that its description or impact on the customer's current status is more important, and vice versa.

[0103] Specifically, the real-time tag weight distribution is normalized, and the weight of each tag is mapped to a unified scale (such as the interval [0,1]) and the sum is made 1 to obtain standardized weight data; based on the standardized weight data, the weight ratio of each tag is calculated (that is, the ratio of the standardized weight of a single tag to the total weight); the tags in the target tag set are prioritized according to the weight ratio to form an importance sequence from high to low; according to the sorting results, the tag order is adjusted, the weight value is updated, and the tags are dynamically replaced or added in combination with the latest behavioral data, and finally a tag set that reflects customer characteristics in real time is generated.

[0104] In an embodiment of the present invention, the real-time portrait refers to the integration of the target customer's current behavioral characteristics and attribute information (such as consumption frequency, consumption potential, etc., weighted data after standardized processing) based on the weight ratio and priority ranking of each tag in the real-time tag set, and dynamically presenting the customer's current status in a visual or structured form, intuitively reflecting the customer's performance and importance in various dimensions.

[0105] In detail, the specific information of each tag in the real-time tag set is extracted (such as tag name, standardized weight data, weight ratio, priority ranking, etc.), and the basic attributes of customers (such as age, gender, and region) are integrated with dynamic behavior data (such as recent consumption amount, browsing time, interaction frequency, and other corresponding tags). According to the tag weight ratio and priority, each dimension of information is assigned a corresponding importance level. Through data visualization tools (such as dashboards, charts) or structured reports, the customer's current characteristics and changing trends in key dimensions such as consumption capacity, interest preferences, and activity are dynamically presented to form a comprehensive and real-time description of customer status.

[0106] Specifically, after generating a real-time profile of a target customer, a streaming computing engine (such as Flink) is used to continuously capture and process new real-time behavioral events (such as instant browsing, transactions, and comments) generated by the customer. When new data triggers weight update conditions, the label weights are automatically recalculated and the real-time label set is updated, thereby achieving dynamic refresh of the real-time profile. At the same time, the profile data is presented in the form of intuitive charts and tag clouds through visualization tools, and a profile query API interface is provided to facilitate business system calls. A label weight explanation diagram is also generated based on algorithms such as decision trees, and the generation path of each label is output (such as "high consumption potential = browsing luxury goods >10 times in the past 30 days + monthly income >50,000"), clearly demonstrating the logic of profile construction and helping business personnel understand the basis for the formation of customer characteristics to support decision-making applications such as precision marketing and customer segmentation.

[0107] Furthermore, a real-time portrait of the target customer is generated based on the real-time tag set. First, various types of information in the real-time tag set are integrated, and the current characteristics of the customer are presented according to the weight ratio and priority. Then, with the help of the streaming computing engine, new behavioral events are processed in real time, and the tag set and portrait are dynamically updated. At the same time, auxiliary functions such as visualization, API interface and tag weight explanation diagram are provided, and the tag generation path is output. Through the above operations, the customer status can be portrayed in real time, accurately and intuitively, which not only improves the timeliness and accuracy of the customer portrait, but also reduces the cost of manual analysis, enhances the interpretability of the data, and helps business personnel quickly understand customer characteristics, providing strong support for decisions such as precision marketing, product optimization, and customer segmentation. In addition, in the entire data processing process, privacy protection mechanisms such as homomorphic encryption and differential privacy are adopted in federated learning to prevent the leakage of original data and ensure data privacy security.

[0108] As can be seen, in the above solution, through multi-source data collection and privacy desensitization, multimodal feature extraction, dynamic labeling system and federated learning modeling, unified mapping of structured and unstructured data is achieved, labels are dynamically updated based on time decay, data privacy is protected by homomorphic encryption, and an interpretable label generation path is output. This invention can increase the accuracy of customer profiling to over 90%, control label update latency to within 1 second, reduce the risk of data privacy leakage by 95%, increase marketing campaign conversion rate by 30%, reduce manual analysis costs by 50%, effectively drive refined enterprise operations, and balance service and data security.

[0109] It should be understood that the order of execution of the steps in the above embodiments does not necessarily mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0110] In one embodiment, a device for generating a customer profile based on multi-source data is provided. The device for generating a customer profile based on multi-source data corresponds to the method for generating a customer profile based on multi-source data in the above embodiment. Figure 5 As shown, the customer profile generation device 100 based on multi-source data includes a target customer data acquisition module 101, a data privacy desensitization module 102, a feature fusion module 103, a real-time tag weight distribution analysis 104 and a real-time profile generation module 105. The functional modules are described in detail as follows:

[0111] The target customer data acquisition module 101 is used to acquire the structured data and unstructured data of the target customer from the target data source;

[0112] The data privacy desensitization module 102 is used to perform privacy desensitization processing on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data;

[0113] The feature fusion module 103 is used to extract the multimodal features and time series features of the structured desensitized data and the unstructured desensitized data respectively, and perform feature fusion on the multimodal features and the time series features to obtain target fusion features;

[0114] A real-time tag weight distribution analysis module 104 is configured to construct a target tag set for the target customer based on the target fusion feature, and analyze the real-time tag weight distribution of the target tag set using a preset federated learning model;

[0115] The real-time portrait generation module 105 is used to update the target tag set in real time according to the real-time tag weight distribution, generate a real-time tag set, and generate a real-time portrait of the target customer according to the real-time tag set.

[0116] In one embodiment, the feature fusion module 103, when extracting the multimodal features and temporal features of the structured desensitized data and the unstructured desensitized data, is configured to:

[0117] Performing text analysis on the structured desensitized data and the unstructured desensitized data respectively to obtain text features;

[0118] Respectively identifying image features of the structured desensitized data and the unstructured desensitized data;

[0119] Determining multimodal features of the structured desensitized data and the unstructured desensitized data based on the text features and the image features;

[0120] Extracting timestamps from the structured desensitized data and the unstructured desensitized data respectively to obtain a data time series;

[0121] The time series features corresponding to the structured desensitized data and the unstructured desensitized data are generated according to the data time series.

[0122] In one embodiment, the feature fusion module 103, when generating the time series features corresponding to the structured desensitized data and the unstructured desensitized data according to the data time series, is further configured to:

[0123] Determining the behavioral frequencies and periodic patterns corresponding to the structured desensitized data and the unstructured desensitized data according to the data time series;

[0124] Extracting features of the behavior frequency and the periodic pattern to obtain a time series feature index;

[0125] The time series features corresponding to the structured desensitized data and the unstructured desensitized data are determined according to the time series feature indicators.

[0126] In one embodiment, the feature fusion module 103, when performing feature fusion on the multimodal feature and the temporal feature to obtain a target fusion feature, is further configured to:

[0127] performing feature standardization processing on the multimodal features and the time series features respectively to obtain standard multimodal features and standard time series features;

[0128] Performing feature mapping on the standard multimodal features and the standard time series features to obtain a mapping node;

[0129] Performing relationship analysis on the mapping nodes to obtain correlation features between the standard multimodal features and the standard time series features;

[0130] The standard multimodal features and the standard time series features are fused according to the correlation features to obtain target fusion features.

[0131] In one embodiment, the real-time tag weight distribution analysis module 104, when executing the process of constructing the target tag set for the target customer based on the target fusion feature, is configured to:

[0132] Performing feature classification on the target fusion features to obtain a feature category set;

[0133] Identify the label type corresponding to each feature in the feature category set;

[0134] Mapping the label types according to a preset label mapping rule to obtain an initial label set;

[0135] A target tag set is determined according to the initial tag set.

[0136] In one embodiment, the real-time tag weight distribution analysis module 104, when analyzing the real-time tag weight distribution of the target tag set using a preset federated learning model, is configured to:

[0137] Using a preset federated learning model to homomorphically encrypt the label data in the label set to obtain a label encryption set;

[0138] Performing differential processing on the encrypted label set to obtain a label differential set;

[0139] Extracting an initial label weight for each label in the label difference set;

[0140] The initial tag weight is dynamically adjusted according to a preset time decay factor to obtain a real-time tag weight distribution.

[0141] In one embodiment, the real-time portrait generation module 105, when performing real-time updating of the target tag set according to the real-time tag weight distribution to generate the real-time tag set, is configured to:

[0142] Normalizing the real-time tag weight distribution to obtain standardized weight data;

[0143] Identifying the weight ratio of each tag in the target tag set according to the standardized weight data;

[0144] Prioritizing the tags in the target tag set using the weight proportions to obtain a ranking result;

[0145] The target tag set is updated in real time according to the sorting result to generate a real-time tag set.

[0146] This invention provides a customer profile generation device based on multi-source data. Through multi-source data collection and privacy desensitization, multimodal feature extraction, a dynamic labeling system, and federated learning modeling, it achieves unified mapping of structured and unstructured data, dynamically updates labels based on time decay, utilizes homomorphic encryption to protect data privacy, and outputs an interpretable label generation path. This invention can increase customer profile accuracy to over 90%, control label update latency to within 1 second, reduce the risk of data privacy leaks by 95%, increase marketing campaign conversion rates by 30%, and reduce manual analysis costs by 50%, effectively driving refined enterprise operations and balancing service and data security.

[0147] For the specific definition of the customer portrait generation device based on multi-source data, please refer to the definition of the customer portrait generation method based on multi-source data above, which will not be repeated here. The various modules in the above-mentioned customer portrait generation device based on multi-source data can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0148] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the service side of a customer portrait generation method based on multi-source data.

[0149] In one embodiment, a computer device is provided. The computer device may be a client, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, memory, network interface, display screen and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps on the client side of a customer profile generation method based on multi-source data.

[0150] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0151] Obtain structured and unstructured data of target customers from target data sources;

[0152] Performing privacy desensitization processing on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data;

[0153] Extracting multimodal features and time series features of the structured desensitized data and the unstructured desensitized data respectively, and fusing the multimodal features and the time series features to obtain target fusion features;

[0154] Constructing a target tag set for the target customer based on the target fusion features, and analyzing the real-time tag weight distribution of the target tag set using a preset federated learning model;

[0155] The target tag set is updated in real time according to the real-time tag weight distribution to generate a real-time tag set, and a real-time portrait of the target customer is generated according to the real-time tag set.

[0156] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0157] Obtain structured and unstructured data of target customers from target data sources;

[0158] Performing privacy desensitization processing on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data;

[0159] Extracting multimodal features and time series features of the structured desensitized data and the unstructured desensitized data respectively, and fusing the multimodal features and the time series features to obtain target fusion features;

[0160] Constructing a target tag set for the target customer based on the target fusion features, and analyzing the real-time tag weight distribution of the target tag set using a preset federated learning model;

[0161] The target tag set is updated in real time according to the real-time tag weight distribution to generate a real-time tag set, and a real-time portrait of the target customer is generated according to the real-time tag set.

[0162] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0163] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0164] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0165] It should be noted that if software tools or components other than those of our company appear in the embodiments of this application, they are only used for illustration and do not represent actual use.

[0166] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A customer profile generation method based on multi-source data, characterized in that: include: Obtain structured and unstructured data of target customers from target data sources; Performing privacy desensitization processing on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data; Extracting multimodal features and time series features of the structured desensitized data and the unstructured desensitized data respectively, and fusing the multimodal features and the time series features to obtain target fusion features; Constructing a target tag set for the target customer based on the target fusion features, and analyzing the real-time tag weight distribution of the target tag set using a preset federated learning model; The target tag set is updated in real time according to the real-time tag weight distribution to generate a real-time tag set, and a real-time portrait of the target customer is generated according to the real-time tag set.

2. The method for generating a customer profile based on multi-source data according to claim 1, wherein: The extracting of multimodal features and temporal features of the structured desensitized data and the unstructured desensitized data respectively includes: Performing text analysis on the structured desensitized data and the unstructured desensitized data respectively to obtain text features; Respectively identifying image features of the structured desensitized data and the unstructured desensitized data; Determining multimodal features of the structured desensitized data and the unstructured desensitized data based on the text features and the image features; Extracting timestamps from the structured desensitized data and the unstructured desensitized data respectively to obtain a data time series; The time series features corresponding to the structured desensitized data and the unstructured desensitized data are generated according to the data time series.

3. The method for generating a customer profile based on multi-source data according to claim 2, wherein: Generating the time series features corresponding to the structured desensitized data and the unstructured desensitized data according to the data time series includes: Determining the behavioral frequencies and periodic patterns corresponding to the structured desensitized data and the unstructured desensitized data according to the data time series; Extracting features of the behavior frequency and the periodic pattern to obtain a time series feature index; The time series features corresponding to the structured desensitized data and the unstructured desensitized data are determined according to the time series feature indicators.

4. The method for generating a customer profile based on multi-source data according to claim 1, wherein: The step of fusing the multimodal features and the time series features to obtain target fusion features includes: performing feature standardization processing on the multimodal features and the time series features respectively to obtain standard multimodal features and standard time series features; Performing feature mapping on the standard multimodal features and the standard time series features to obtain a mapping node; Performing relationship analysis on the mapping nodes to obtain correlation features between the standard multimodal features and the standard time series features; The standard multimodal features and the standard time series features are fused according to the correlation features to obtain target fusion features.

5. The method for generating a customer profile based on multi-source data according to claim 1, wherein: The step of constructing a target tag set for the target customer based on the target fusion feature includes: Performing feature classification on the target fusion features to obtain a feature category set; Identify the label type corresponding to each feature in the feature category set; Mapping the label types according to a preset label mapping rule to obtain an initial label set; A target tag set is determined according to the initial tag set.

6. The method for generating a customer profile based on multi-source data according to claim 1, wherein: The using of a preset federated learning model to analyze the real-time tag weight distribution of the target tag set includes: Using a preset federated learning model to homomorphically encrypt the label data in the label set to obtain a label encryption set; Performing differential processing on the encrypted label set to obtain a label differential set; Extracting an initial label weight for each label in the label difference set; The initial tag weight is dynamically adjusted according to a preset time decay factor to obtain a real-time tag weight distribution.

7. The method for generating a customer profile based on multi-source data according to claim 1, wherein: The step of updating the target tag set in real time according to the real-time tag weight distribution to generate a real-time tag set includes: Normalizing the real-time tag weight distribution to obtain standardized weight data; Identifying the weight ratio of each tag in the target tag set according to the standardized weight data; Prioritizing the tags in the target tag set using the weight proportions to obtain a ranking result; The target tag set is updated in real time according to the sorting result to generate a real-time tag set.

8. A customer profile generation device based on multi-source data, characterized in that: include: A target customer data acquisition module is used to acquire the structured and unstructured data of target customers from a target data source; A data privacy desensitization module is used to perform privacy desensitization processing on the structured data and the unstructured data respectively to obtain structured desensitized data and unstructured desensitized data; A feature fusion module is used to extract multimodal features and time series features of the structured desensitized data and the unstructured desensitized data, respectively, and fuse the multimodal features and the time series features to obtain target fusion features; A real-time tag weight distribution analysis module is used to construct a target tag set for the target customer based on the target fusion feature, and analyze the real-time tag weight distribution of the target tag set using a preset federated learning model; The real-time portrait generation module is used to update the target tag set in real time according to the real-time tag weight distribution, generate a real-time tag set, and generate a real-time portrait of the target customer according to the real-time tag set.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for generating a customer profile based on multi-source data as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for generating a customer profile based on multi-source data as described in any one of claims 1 to 7 is implemented.

Images