User behavior analysis method based on multi-source information

Abstract

The present invention discloses a user behavior analysis method based on multi-source information, and obtains a user portrait based on the user's consumption behavior. The method pre-processes the user's consumption record and payment data, and extracts a user behavior feature vector, wherein the feature vector includes consumption amount, consumption frequency, payment method and consumption category; the feature vector is encoded based on a preset neural network model to obtain a standardized representation of the user behavior; based on the standardized representation of the user behavior, a cluster analysis is performed on the user through a preset genetic algorithm to obtain a user portrait; the present invention significantly improves the accuracy and comprehensiveness of the user portrait, and provides strong support for the personalized recommendation and precision marketing of the enterprise.

Description

Technical Field

[0001] The present invention relates to the technical field of data mining, and in particular to a user behavior analysis method based on multi-source information. Background Art

[0002] User behavior refers to the various interactive behaviors generated by users in the process of using products or services, such as consumption records, payment methods, consumption categories, etc. By analyzing user behavior, a user portrait can be generated, which is an abstract model that describes user characteristics and behavior patterns, usually including multiple dimensions such as consumption ability, consumption preferences, consumption behavior and user value. User portraits are an important basis for enterprises to make personalized recommendations, formulate marketing strategies and optimize customer service.

[0003] Most existing user behavior analysis methods rely on traditional statistical methods and simple machine learning models, such as clustering algorithms (such as K-means) and decision trees. Although these methods are simple and easy to use, they often face many challenges when dealing with large-scale, high-dimensional, nonlinear and time-series data. For example, the K-means algorithm is very sensitive to the selection of the initial center point, easily falls into the local optimal solution, and cannot handle noise and outliers well. In addition, traditional machine learning models often require manual feature design, which is not only time-consuming and labor-intensive, but also may miss important user behavior patterns. Therefore, existing technologies have certain limitations in user behavior analysis and are difficult to meet the growing needs of data analysis. Summary of the invention

[0004] The purpose of the present invention is to provide a user behavior analysis method based on multi-source information, pre-process consumption records and payment data, and use neural network models and genetic algorithms to perform standardized representation and cluster analysis of user behavior, thereby achieving the generation and update of user portraits.

[0005] In order to achieve the above-mentioned purpose of the present invention, the present invention adopts the following technical solutions:

[0006] A user behavior analysis method based on multi-source information, obtaining a user portrait based on the user consumption behavior, the method comprising:

[0007] Step S1, pre-processing the user's consumption records and payment data to extract the user behavior feature vector, which includes consumption amount, consumption frequency, payment method and consumption category.

[0008] Step S2, encoding the feature vector based on a preset neural network model to obtain a standardized representation of the user behavior.

[0009] The neural network model includes: an input layer, an embedding layer, two hidden layers, and an output layer.

[0010] The input layer is used to receive the user behavior feature vector; the embedding layer is used to convert the feature vector into a The embedding vector The two hidden layers include a first hidden layer and a second hidden layer, the first hidden layer is layer, used to process time series features, the second hidden layer is a fully connected layer; the output layer is used to generate a standardized representation of user behavior .

[0011] Step S3, based on the standardized representation of user behavior, cluster analysis is performed on the users through a preset genetic algorithm to obtain user portraits.

[0012] Furthermore, the input layer is used to receive the user behavior feature vector. The embedding layer is used to convert the feature vector into a dimension The embedding vector , specifically:

[0013] The input layer receives the user behavior feature vector , whose dimensions are ,in is the feature dimension;

[0014] The embedding layer is constructed by the projection matrix Convert the input vector into an embedding vector : , where the projection matrix The optimization objective function is: , the constraints are: ;in, It means finding the parameters that minimize the objective function. , Represents the feature vectors of different samples, is the similarity matrix, representing the sample and samples The similarity between is a diagonal matrix and , indicating the sample The sum of the similarities with all other samples; I is the unit matrix, which is used to maintain the orthogonality of the mapped features; Represents the square of the Euclidean distance between two samples after mapping;

[0015] Bias Vector By minimizing the reconstruction error: ,in, Represents the reconstruction error, which is the difference between the mapped data and the original data.

[0016] Furthermore, the embedding vector As the input sequence of the first hidden layer RNN, the calculation formula of the first hidden layer RNN is:

[0017] ,in, is the input at time t, indicating that from the embedding vector at time step t The corresponding features extracted from ; is the hidden state at time t, and are the input transformation matrix and the state transfer matrix respectively; the bias term is divided into input bias and state bias :

[0018] , ,in, is the weight matrix that transforms the input into the input bias, is the weight matrix that converts the previous state into the state bias;

[0019] The output of the second hidden layer The calculation formula is: ,in, is the weight matrix, and the bias term contains the static bias and dynamic bias :

[0020] , ;in, is a scalar parameter for static bias, which is multiplied by the all-1 vector to produce a fixed bias. is a vector of all 1s, is a dynamic bias transformation matrix that generates adaptive biases based on hidden states;

[0021] The output layer generates a normalized representation of user behavior :

[0022] ,in, is the weight matrix that transforms the second layer output to the final output space, is the class bias, For attention bias:

[0023] , ;in, is a learnable scalar parameter, is the attention weight matrix.

[0024] Furthermore, based on the preset neural network model, 75% of the samples are selected for parameter training, and 25% of the samples are selected for verification. After verifying the neural network model, the neural network parameters are optimized based on the back propagation algorithm and momentum term optimization:

[0025] Step S21, calculating the cross entropy loss of the neural network, adding the L2 regularization term to obtain the total loss function;

[0026] ,in, is the sample size, is the number of categories, is the true label, is the predicted value, is the regularization coefficient, is the set of all weight parameters;

[0027] Step S22, calculating the gradient of the loss function with respect to each parameter, and updating the network parameters based on the back propagation algorithm;

[0028] ,in, is the weight parameter of the tth iteration, is the learning rate;

[0029] Step S23, introducing a momentum term to optimize parameter update and obtain an updated neural network model;

[0030] , ;

[0031] in, is the momentum term.

[0032] Furthermore, the clustering analysis of users by using a preset genetic algorithm includes:

[0033] Step S31: Standardize the user behavior Converted into chromosome sequence for encoding. During the encoding process, each gene position corresponds to a user, and its value represents the cluster category number to which the user belongs;

[0034] Step S32, calculating the fitness by considering both the intra-class compactness and the inter-class separation;

[0035] Fitness The calculation formula is:

[0036] ,in, is the number of categories, For the The sample set of the class, For sample and The distance is the weight coefficient;

[0037] Step S33, starting an evolution operation based on the fitness;

[0038] Using the roulette wheel selection method, the selection probability is: ; For the The fitness of an individual, is the number of individuals in the population;

[0039] Using two-point crossover, the crossover probability is ;

[0040] Adaptive mutation is used, and the mutation probability is: ,in, is the initial mutation probability, is the maximum fitness in the current population.

[0041] Furthermore, the termination condition of the genetic algorithm is: The change rate of the optimal fitness value of each generation is less than the preset threshold Terminates when , or when the number of iterations reaches the preset maximum number of generations Terminated.

[0042] Furthermore, the user portrait includes four dimensions: consumption capacity dimension, consumption preference dimension, consumption behavior dimension and user value dimension.

[0043] The consumption capacity dimensions include: average monthly consumption amount range, maximum single consumption amount, and consumption frequency;

[0044] The consumption preference dimensions include: commonly used payment methods, most frequently purchased categories, and category preference intensity;

[0045] The consumption behavior dimensions include: consumption time distribution, consumption periodicity, and promotion sensitivity;

[0046] The user value dimensions include: customer life cycle value, loyalty index, and churn risk level.

[0047] Furthermore, the user portrait is dynamically updated based on the user's real-time consumption behavior, which includes: adding new consumption records, changing payment methods, and expanding consumer product categories.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] By combining deep learning and genetic algorithms, the present invention realizes efficient analysis and clustering of user behaviors, significantly improving the accuracy and comprehensiveness of user portraits. It can not only automatically extract and learn user behavior characteristics, but also effectively process large-scale, high-dimensional and time series data, overcoming the limitations of traditional methods. It encodes user behavior characteristics through a deep neural network model and uses a genetic algorithm for clustering analysis to ultimately generate high-quality user portraits, providing strong support for the company's personalized recommendations and precision marketing. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of a method for analyzing user behavior based on multi-source information according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] All actions of obtaining information or data in the present invention are carried out in compliance with the corresponding data protection laws and policies of the country where the user is located and with the authorization of the corresponding user.

[0053] The present invention analyzes user behavior by integrating multi-source information, where the multi-source information includes but is not limited to:

[0054] Transaction data: users’ shopping records, payment details, return and exchange information, etc.

[0055] Behavioral data: user browsing history, favorites, shopping cart data, evaluation feedback, etc.

[0056] User attribute data: basic demographic characteristics, membership level, points, etc.;

[0057] External data: social media data, location information, weather data, etc.

[0058] The integration of these multi-source information provides a basis for a comprehensive understanding of user behavior. For example, an e-commerce platform can analyze users' consumption behavior in different scenarios and find that users tend to browse food products during lunch breaks on weekdays, while they pay more attention to home furnishings and clothing products on weekends. Combined with weather information, it can also be found that users are more inclined to shop online on rainy days, and the number of takeout orders increases significantly.

[0059] like Figure 1FIG. 1 is a flow chart of a method for analyzing user behavior based on multi-source information according to an embodiment of the present invention. A user portrait is obtained based on the user consumption behavior. The method includes:

[0060] Step S1, pre-processing the user's consumption records and payment data to extract the user behavior feature vector, which includes consumption amount, consumption frequency, payment method and consumption category.

[0061] This method first comprehensively collects and preprocesses user behavior data. Taking a comprehensive e-commerce platform as an example, the system collects the following data:

[0062] When preprocessing the user's consumption records and payment data, it specifically includes: data cleaning, processing outliers: for example, a single transaction amount that exceeds the user's highest historical consumption amount by more than 3 times is marked as a suspected anomaly, supplementing missing values: for missing consumption categories, intelligent classification is performed based on the product name and description, and format unification: data from different payment channels are unified into a standard format;

[0063] Consumption amount: calculate the average daily and monthly consumption amount and construct the consumption amount distribution characteristics;

[0064] Consumption frequency: Count the purchase frequency on a daily, weekly, and monthly basis to identify peak consumption periods;

[0065] Payment methods: Analyze the usage ratio of various payment methods and calculate the diversity index of payment methods;

[0066] Consumer categories: collect statistics on the purchase frequency and amount of each category, and build a category preference matrix;

[0067] For example, the original data of a user named Xiao Wang shows that there were 47 transaction records in the past three months, with an average monthly consumption of 5,200 yuan. The payment method distribution is as follows: Alipay accounts for 60%, WeChat Pay accounts for 30%, and bank cards account for 10%. The main consumer categories are: beauty and skin care (35%), clothing (25%), digital (20%), and others (20%).

[0068] Step S2, encoding the feature vector based on a preset neural network model to obtain a standardized representation of the user behavior.

[0069] The neural network model includes: an input layer, an embedding layer, two hidden layers, and an output layer.

[0070] The input layer is used to receive the user behavior feature vector; the embedding layer is used to convert the feature vector into a The embedding vector The two hidden layers include a first hidden layer and a second hidden layer, the first hidden layer is layer, used to process time series features, the second hidden layer is a fully connected layer; the output layer is used to generate a standardized representation of user behavior .

[0071] The system receives the preprocessed user behavior feature vector, for example: user ID: 10086, feature vector: [5200, 47, 0.6, 0.3, 0.1, 0.35, 0.25, 0.2, 0.2]. The vector here contains information such as average monthly consumption, number of transactions, proportion of each payment method, and proportion of each category.

[0072] The projection matrix converts the high-dimensional sparse original features into low-dimensional dense representations. For example, the above 9-dimensional vector is mapped to a 128-dimensional embedding vector, so that users with similar behaviors are closer in the embedding space. For the processing of time series data, such as analyzing the changes in user consumption patterns for three consecutive months: January: mainly consuming beauty products, February: increased clothing consumption, March: began to pay attention to digital products, the RNN layer can capture the evolution of this consumption habit.

[0073] Through multiple layers of transformation, we can eventually get a standardized representation of user behavior; for example, users may be classified into composite labels such as "high consumption, love of beauty and fashion, digital enthusiast".

[0074] The input layer is used to receive the user behavior feature vector, and the embedding layer is used to convert the feature vector into a The embedding vector , specifically:

[0075] The input layer receives the user behavior feature vector , whose dimensions are ,in is the feature dimension.

[0076] The embedding layer is constructed by the projection matrix Convert the input vector into an embedding vector : , where the projection matrix The optimization objective function is: , the constraints are: ;in, It means finding the parameters that minimize the objective function. , Represents the feature vectors of different samples, is the similarity matrix, representing the sample and samples The similarity between is a diagonal matrix and , indicating the sample The sum of the similarities with all other samples; I is the unit matrix, which is used to maintain the orthogonality of the mapped features; Represents the square of the Euclidean distance between two samples after mapping.

[0077] Bias Vector By minimizing the reconstruction error: ,in, Represents the reconstruction error, which is the difference between the mapped data and the original data.

[0078] Embedding vector As the input sequence of the first hidden layer RNN, the calculation formula of the first hidden layer RNN is:

[0079] ,in, is the input at time t, indicating that from the embedding vector at time step t The corresponding features extracted from ; is the hidden state at time t, and are the input transformation matrix and the state transfer matrix respectively; the bias term is divided into input bias and state bias :

[0080] , ,in, is the weight matrix that transforms the input into the input bias, is the weight matrix that transforms the previous state into the state bias.

[0081] The output of the second hidden layer The calculation formula is: ,in, is the weight matrix, and the bias term contains the static bias and dynamic bias :

[0082] , ;in, is a scalar parameter for static bias, which is multiplied by the all-1 vector to produce a fixed bias. is a vector of all 1s, is a dynamic bias transformation matrix that generates adaptive biases based on hidden states;

[0083] The output layer generates a normalized representation of user behavior :

[0084] ,in, is the weight matrix that transforms the second layer output to the final output space, is the class bias, For attention bias:

[0085] , ;in, is a learnable scalar parameter, is the attention weight matrix.

[0086] Based on the preset neural network model, 75% of the samples are selected for parameter training and 25% of the samples are selected for verification. After verifying the neural network model, the neural network parameters are optimized based on the back propagation algorithm and momentum term optimization:

[0087] Step S21, calculate the cross entropy loss of the neural network, and add the L2 regularization term to obtain the total loss function.

[0088] ,in, is the sample size, is the number of categories, is the true label, is the predicted value, is the regularization coefficient, is the set of all weight parameters;

[0089] Step S22, calculate the gradient of the loss function with respect to each parameter, and update the network parameters based on the back propagation algorithm.

[0090] ,in, is the weight parameter of the tth iteration, is the learning rate;

[0091] Step S23, introducing the momentum term to optimize parameter update and obtain an updated neural network model.

[0092] , ;

[0093] in, is the momentum term.

[0094] Step S3, based on the standardized representation of user behavior, cluster analysis is performed on the users through a preset genetic algorithm to obtain user portraits.

[0095] The clustering analysis of users by using a preset genetic algorithm includes:

[0096] Step S31: Standardize the user behavior Converted into chromosome sequence for encoding. During the encoding process, each gene position corresponds to a user, and its value represents the cluster category number to which the user belongs.

[0097] Step S32, calculating the fitness by considering both the intra-class compactness and the inter-class separation.

[0098] Fitness The calculation formula is:

[0099] ,in, is the number of categories, For the The sample set of the class, For sample and The distance is the weight coefficient.

[0100] Step S33, starting the evolution operation based on the fitness.

[0101] Using the roulette wheel selection method, the selection probability is: ; For the The fitness of an individual, is the number of individuals in the population;

[0102] Using two-point crossover, the crossover probability is ;

[0103] Adaptive mutation is used, and the mutation probability is: ,in, is the initial mutation probability, is the maximum fitness in the current population.

[0104] The termination condition of the genetic algorithm is: The change rate of the optimal fitness value of each generation is less than the preset threshold Terminates when , or when the number of iterations reaches the preset maximum number of generations Terminated.

[0105] In the coding design, it is assumed that the system needs to divide users into 5 categories, and the gene of each user can be represented by an integer from 0 to 4. For example: User A: 2 (indicates belonging to the third category), User B: 0 (indicates belonging to the first category), User C: 4 (indicates belonging to the fifth category). Considering the actual business needs, a comprehensive scoring mechanism is designed: the intra-class similarity weight α = 0.6, the inter-class difference weight (1-α) = 0.4, such a setting ensures that the clustering results maintain both the similarity of users within the class and the difference between different categories. In actual operation, the following parameters are set: population size: 100, crossover probability: 0.8, initial mutation probability: 0.1, maximum number of generations: 200, convergence threshold: 0.001; with such parameter settings, the algorithm can usually reach convergence in about 100 generations.

[0106] The user portrait includes four dimensions: consumption capacity dimension, consumption preference dimension, consumption behavior dimension and user value dimension; the four dimensions of the user portrait cover multiple aspects of the user, which helps enterprises better understand and manage users.

[0107] Consumption capacity dimension: including the average monthly consumption amount range, the highest single consumption amount, and consumption frequency. These indicators can help companies identify users with high consumption capacity and formulate targeted marketing strategies.

[0108] Consumption preference dimension: including common payment methods, most frequently purchased categories, and category preference strength. For example, some users may prefer to pay with credit cards, or prefer to buy electronic products. By understanding users' consumption preferences, companies can launch products and services that meet user needs.

[0109] Consumption behavior dimension: including consumption time distribution, consumption periodicity, and promotion sensitivity. For example, some users may shop more on weekends or be particularly sensitive to discounts and promotions. This information helps companies arrange promotional activities and optimize inventory management.

[0110] User value dimensions: including customer lifetime value, loyalty index, and churn risk level. For example, by analyzing the user's purchase history and feedback, you can assess the user's loyalty and churn risk, and take measures to improve user retention.

[0111] Dynamically update user portraits based on real-time consumer behavior, including: new consumer records, changes in payment methods, and expansion of consumer categories. User behavior is constantly changing, so user portraits also need to be updated in real time to reflect the latest situation. For example, if a user suddenly increases the frequency of consumption or tries a new payment method, this may indicate that the user's living conditions have changed or that they have developed an interest in new products. By monitoring user consumption behavior in real time and updating user portraits in a timely manner, companies can better track user dynamics and provide personalized services and products, thereby improving user experience and satisfaction.

[0112] The following are some examples of specific application scenarios for user portraits:

[0113] From the consumption ability dimension, the portrait of a user named Mr. Zhang shows: average monthly consumption range: 8,000-10,000 yuan, highest single consumption: 15,000 yuan, average monthly consumption frequency: 12 times; based on these data, the system classifies him as a "high-value customer" and can give priority to pushing high-end product information.

[0114] In terms of consumption preference, Ms. Li’s preference profile is as follows: commonly used payment method: credit card (80%), most commonly purchased categories: beauty and skin care, luxury goods, category preference intensity: beauty (0.8), luxury goods (0.7). Based on this, the system pushes information about new cosmetics and limited edition luxury goods to her.

[0115] In terms of consumption behavior, Mr. Wang’s behavior characteristics are as follows: consumption period: weekday evenings (accounting for 60%), consumption cycle: concentrated shopping at the beginning of each month, promotion sensitivity: high (coupon usage rate 90%). Based on these characteristics, the system pushes coupons to him in the evenings at the beginning of each month.

[0116] In terms of user value, Ms. Zhao's value assessment is as follows: customer lifetime value: estimated to be 50,000 yuan in the next three years, loyalty index: 85 (high loyalty), churn risk: low (5%). Based on these indicators, she will be included in the VIP customer retention plan.

[0117] The specific implementation of the real-time update mechanism includes some trigger conditions, such as: the single-day consumption amount exceeds the historical high, the new payment method is added, and the new category of goods is purchased for the first time. For example: the update process of Xiao Chen's portrait: the original portrait: medium consumption ability, preference for digital products, trigger event: continuous purchase of multiple luxury goods, update result: increase the consumption ability level, add luxury preference label; for example: the system detects changes in user consumption behavior: the recent consumption frequency has dropped by 50%, the browsing time has decreased, and the shopping cart items have not been settled for a long time. The system will increase the risk level of loss and trigger the retention strategy. Through such a real-time update mechanism, it is ensured that the user portrait can promptly reflect the changes in user behavior and provide more accurate marketing services.

[0118] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A user behavior analysis method based on multi-source information, which obtains a user profile based on user consumption behavior, characterized in that: The method comprises: Step S1, pre-processing the user's consumption records and payment data to extract the user behavior feature vector, which includes consumption amount, consumption frequency, payment method and consumption category; Step S2, encoding the feature vector based on a preset neural network model to obtain a standardized representation of the user behavior; The neural network model includes: an input layer, an embedding layer, two hidden layers, and an output layer; The input layer is used to receive the user behavior feature vector; the embedding layer is used to convert the feature vector into a The embedding vector The two hidden layers include a first hidden layer and a second hidden layer, the first hidden layer is layer, used to process time series features, the second hidden layer is a fully connected layer; the output layer is used to generate a standardized representation of user behavior ; The input layer is used to receive the user behavior feature vector, and the embedding layer is used to convert the feature vector into a The embedding vector , specifically: The input layer receives the user behavior feature vector , whose dimensions are ,in is the feature dimension; The embedding layer is constructed by the projection matrix Convert the input vector into an embedding vector : , where the projection matrix The optimization objective function is: , the constraints are: ;in, It means finding the parameters that minimize the objective function. , Represents the feature vectors of different samples, is the similarity matrix, representing the sample and samples The similarity between is a diagonal matrix and , indicating the sample The sum of the similarities with all other samples; I is the unit matrix, which is used to maintain the orthogonality of the mapped features; Represents the square of the Euclidean distance between two samples after mapping; Bias Vector By minimizing the reconstruction error: ,in, represents the reconstruction error, that is, the difference between the mapped data and the original data; Step S3, based on the standardized representation of user behavior, cluster analysis is performed on the users through a preset genetic algorithm to obtain user portraits; The clustering analysis of users by using a preset genetic algorithm includes: Step S31: Standardize the user behavior Converted into chromosome sequence for encoding; during the encoding process, each gene position corresponds to a user, and its value represents the cluster category number to which the user belongs; Step S32, calculating the fitness by considering both the intra-class compactness and the inter-class separation; Fitness The calculation formula is: ,in, is the number of categories, For the The sample set of the class, For sample and The distance is the weight coefficient; Step S33, starting an evolution operation based on the fitness; Using the roulette wheel selection method, the selection probability is: ; For the The fitness of an individual, is the number of individuals in the population; Using two-point crossover, the crossover probability is ; Adaptive mutation is used, and the mutation probability is: ,in, is the initial mutation probability, is the maximum fitness in the current population.

2. The user behavior analysis method based on multi-source information according to claim 1 is characterized in that: Embedding vector As the input sequence of the first hidden layer RNN, the calculation formula of the first hidden layer RNN is: ,in, is the input at time t, indicating that from the embedding vector at time step t The corresponding features extracted from ; is the hidden state at time t, and are the input transformation matrix and the state transfer matrix respectively; the bias term is divided into input bias and state bias : , ,in, is the weight matrix that transforms the input into the input bias, is the weight matrix that converts the previous state into the state bias; The output of the second hidden layer The calculation formula is: ,in, is the weight matrix, and the bias term contains the static bias and dynamic bias : , ;in, is a scalar parameter for static bias, which is multiplied by the all-1 vector to produce a fixed bias. is a vector of all 1s, is a dynamic bias transformation matrix that generates adaptive biases based on hidden states; The output layer generates a normalized representation of user behavior : ,in, is the weight matrix that transforms the second layer output to the final output space, is the class bias, For attention bias: , ;in, is a learnable scalar parameter, is the attention weight matrix.

3. The user behavior analysis method based on multi-source information according to claim 2 is characterized in that: Based on the preset neural network model, 75% of the samples are selected for parameter training and 25% of the samples are selected for verification. After verifying the neural network model, the neural network parameters are optimized based on the back propagation algorithm and momentum term optimization: Step S21, calculating the cross entropy loss of the neural network, adding the L2 regularization term to obtain the total loss function; ,in, is the sample size, is the number of categories, is the true label, is the predicted value, is the regularization coefficient, is the set of all weight parameters; Step S22, calculating the gradient of the loss function with respect to each parameter, and updating the network parameters based on the back propagation algorithm; ,in, is the weight parameter of the tth iteration, is the learning rate; Step S23, introducing a momentum term to optimize parameter update and obtain an updated neural network model; ， ； in, is the momentum term.

4. The user behavior analysis method based on multi-source information according to claim 3 is characterized in that: The termination condition of the genetic algorithm is: The change rate of the optimal fitness value of each generation is less than the preset threshold Terminates when , or when the number of iterations reaches the preset maximum number of generations Terminated.

5. The user behavior analysis method based on multi-source information according to claim 4 is characterized in that: The user portrait includes four dimensions: consumption capacity dimension, consumption preference dimension, consumption behavior dimension and user value dimension; The consumption capacity dimensions include: average monthly consumption amount range, maximum single consumption amount, and consumption frequency; The consumption preference dimensions include: commonly used payment methods, most frequently purchased categories, and category preference intensity; The consumption behavior dimensions include: consumption time distribution, consumption periodicity, and promotion sensitivity; The user value dimensions include: customer life cycle value, loyalty index, and churn risk level.

6. The user behavior analysis method based on multi-source information according to claim 5 is characterized in that: The user portrait is dynamically updated based on the user's real-time consumption behavior, which includes: new consumption records, payment method changes, and consumer category expansion.

Images