Customer behavior prediction and intelligent strategy generation method and system based on multi-source data fusion

Abstract

The invention provides a customer behavior prediction and intelligent strategy generation method and system based on multi-source data fusion. According to the method, client data is collected from multiple sources such as a relational database, a log file and a social media platform, and after cleaning and normalization preprocessing, features are fused and weights are determined. Then, a deep neural network prediction model is constructed based on the fused data, and an activation function and a loss function are selected for training according to a prediction target; and finally, generating an intelligent strategy by using a Q-learning algorithm according to a prediction result. The system comprises a data acquisition module, a preprocessing module, a fusion module, a prediction model construction module, a training module and an intelligent strategy generation module, and all the modules cooperate to complete conversion from data to a strategy. Through multi-source data fusion, customer characteristics are comprehensively described, the accuracy of customer behavior prediction is remarkably improved, the generated intelligent strategy is high in pertinence and effectiveness, the market competitiveness of enterprises is effectively enhanced, and the method has extremely high application value in the fields of customer behavior analysis and strategy making.

Description

Technical Field

[0001] The present invention relates to the fields of data analysis, artificial intelligence, and computer applications, and particularly to a method and system for integrating multi-source data to achieve accurate customer behavior prediction and intelligent strategy generation, aiming to solve the problems of inaccurate prediction and lack of pertinence in customer behavior analysis and strategy formulation for enterprises due to scattered data sources and single processing methods. Background Art

[0002] In the era of big data, the customer data owned by enterprises shows explosive growth and extremely wide sources. Online trading platforms record every order information, browsing track, and payment method of customers; offline stores store face-to-face transaction records, membership registration information, and on-site service feedback of customers; social media platforms contain multi-dimensional information such as customer evaluations of brands, sharing behaviors, and interest preferences; customer service feedback data also includes problems encountered by customers, complaint content, and suggestions for product improvement. However, these data have serious dispersion and heterogeneity. The data formats, data structures, and data meanings of different data sources are very different, and traditional data processing methods are difficult to effectively integrate these data. This makes it impossible for enterprises to form a comprehensive and in-depth understanding of customers when analyzing customer behavior, and can only analyze based on one-sided data, resulting in low accuracy of customer behavior prediction. Strategies formulated based on a single data source often lack pertinence in actual applications and cannot meet the urgent needs of enterprises for refined customer operations in the fierce market competition. Summary of the Invention

[0003] Object of the Invention

[0004] The core object of the present invention is to provide a method and system for customer behavior prediction and intelligent strategy generation based on multi-source data fusion. By efficiently integrating multi-source data, mining the potential value behind the data, and constructing an accurate customer behavior prediction model, highly personalized, effective intelligent strategies that meet the actual business needs of enterprises can be generated, comprehensively improving the customer operation efficiency and market competitiveness of enterprises, and helping enterprises gain advantages in the complex and changeable market environment.

[0005] Technical Solution

[0006] 1. Multi-source Data Collection and Preprocessing

[0007] Data collection: Collect data from multiple data sources, including relational databases, log files, social media platform interfaces, etc. Let the data of the th data source collected be where . Taking a relational database as an example, it may involve extracting relevant data from different table structures, such as obtaining transaction amounts and transaction times from an "order table" and obtaining basic customer information from a "user information table". For log files, it is necessary to parse log records in different formats and extract key information, such as extracting the browsed pages and stay durations from the browsing behavior logs of users.

[0008] Data cleaning: Use statistical-based methods to identify and process outliers. For numerical data, utilize the interquartile range to judge outliers, and the formula is: ; ;

[0009] where, is the first quartile, is the third quartile. Data points outside the upper and lower limits are regarded as outliers and are corrected or deleted. In actual operations, for transaction amount data, if the transaction amount of a certain order far exceeds the normal range and is judged as an outlier through the above formula, the data source can be further verified. If it is an input error, it is corrected; if it cannot be verified, the abnormal data point is deleted.

[0010] Data normalization: Normalize data of different scales to make the data comparable. For data x, adopt the min-max normalization method, and the formula is:

[0011] where, and are the minimum and maximum values in the dataset respectively. For example, after normalizing customer purchase amount data, purchase amounts of different magnitudes can be unified into the [0, 1] interval, which is convenient for subsequent model processing.

[0012] 2. Multi-source data fusion

[0013] Feature extraction: Extract feature vectors from each data source . From online transaction data, the purchase frequency , average purchase amount , purchase category preference , etc. of customers can be extracted; from offline store data, the membership level , consumption times , consumption period , etc. of customers can be extracted; from social media data, the sentiment tendency (judged as positive, negative or neutral through text analysis), interaction frequency , etc. of customers can be extracted.

[0014] Feature fusion: Adopt the method of feature fusion to splice the feature vectors of different data sources. Let the feature vectors extracted from the data sources be The fused feature vector is:

[0015] For example, splice the feature vector of online transaction data , the feature vector of offline store data and the feature vector of social media data into a comprehensive feature vector.

[0016] Weight determination: To measure the importance of different data sources, introduce the weight coefficient , and determine the optimal weight through cross-validation. The fused feature vector is expressed as:

[0017] When determining the weights, through multiple cross-validation experiments, calculate indicators such as the prediction accuracy of the model under different weight combinations, and select the weight combination that makes the indicators optimal. For example, determine the weight of online transaction data , the weight of offline store data , the weight of social media data .

[0018] 3. Construction of customer behavior prediction model

[0019] Model architecture: Adopt a deep neural network (DNN) as the prediction model. The model structure includes an input layer, multiple hidden layers, and an output layer. The input layer receives the fused feature vector . The setting of the hidden layer is crucial for the model performance. Multiple hidden layers can automatically learn the complex feature representations of the data.

[0020] Activation function: In the hidden layer, use the rectified linear unit (ReLU) as the activation function. The formula is: where x is the input of the neuron and y is the output. The ReLU function can effectively solve the gradient vanishing problem and accelerate the model convergence speed.

[0021] Hidden layer calculation: Let the weight matrix of the hidden layer be , the bias vector be , and the output of the l-th layer is calculated by the following formula: where . The calculation process of each layer is to perform a linear transformation on the output of the previous layer and then process it through the activation function, continuously extracting the high-level features of the data.

[0022] Output layer setting: According to the prediction target, different activation functions and loss functions are adopted in the output layer. If predicting the probability of customer purchase, the sigmoid activation function is used, and the formula is:

[0023] The binary cross-entropy loss function is used as the loss function, and the formula is:

[0024] Among them, is the number of samples, is the true label, is the predicted value. The sigmoid function maps the output value to the interval [0, 1] to represent the probability, and the binary cross-entropy loss function can effectively measure the difference between the predicted value and the true value.

[0025] 4. Intelligent strategy generation

[0026] Definition of state and action: Based on the prediction results, a reinforcement learning algorithm is used to generate intelligent strategies. Define the state space as the customer's characteristics and prediction results, and the action space as the strategies that the enterprise can take, such as recommending products, sending coupons, pushing personalized advertisements, etc. For example, the state space can be expressed as , where is the fused customer feature vector, is the predicted probability of customer purchase.

[0027] Q-learning algorithm: The Q-learning algorithm is used to learn the optimal strategy, and the update formula for the Q value is: Among them, is the learning rate, is the reward, is the discount factor, is the next state. In practical applications, if the customer purchases the product after the recommendation, a positive reward is given, and the value is updated. Through continuous iterative learning, the model can select the optimal strategy according to different customer states.

[0028] Policy iteration optimization: Through continuous iterative learning, the optimal strategy in different states is obtained, that is, according to the customer's characteristics and predicted behaviors, the strategy that can best improve the enterprise's goals (such as sales volume, customer satisfaction, etc.) is selected. During the iteration process, the learning rate and discount factor are continuously adjusted to balance the exploration and exploitation of the strategy, making the strategy more optimized.

[0029] Innovation Highlights of the Invention

[0030] The invention shows significant innovation in aspects such as data fusion, model construction, and strategy generation. It effectively integrates multi-source data, accurately predicts customer behavior, and generates intelligent strategies, providing strong support for enterprise operations and enhancing market competitiveness.

[0031] 1. Innovation in Multi-source Data Fusion

[0032] Innovation in data source integration: Break down data barriers and integrate multi-source data such as relational databases, log files, and social media platform interfaces. Collect customer transaction, browsing, and social interaction data from different data sources to comprehensively depict the customer portrait, breaking through the limitations of traditional single-data-source analysis and providing a rich data foundation for accurate prediction and strategy formulation. Taking a financial services company as an example, integrating online investment transaction, offline branch consultation, and social media comment data enables the enterprise to understand customers' investment preferences, risk tolerance, and emotional tendencies from multiple dimensions.

[0033] Innovation in fusion methods: Adopt feature splicing and weighted fusion, and determine the weights through cross-validation. For example, based on the contribution of online transactions, offline business outlets, and social media data to customer behavior prediction, determine their weights to be 0.4, 0.3, and 0.3 respectively, effectively integrating information from different data sources. Introduce an adaptive weight adjustment mechanism to recalculate the weights regularly according to data dynamic changes and model feedback, enabling the fused data to more accurately reflect customer characteristics and improving the accuracy of prediction and strategy generation.

[0034] 2. Innovation in Customer Behavior Prediction Model

[0035] Innovation in model architecture: Build a deep neural network prediction model, introducing residual connections and attention mechanisms on the basis of the traditional structure. Residual connections solve the problem of gradient disappearance, enabling the model to learn deeper-level features; the attention mechanism focuses on key features and improves prediction accuracy. For example, in the e-commerce scenario, the model can better capture complex customer purchase behavior patterns, and compared with traditional neural network models, the prediction accuracy is increased by 15% - 25%.

[0036] Innovation in activation function and calculation: Use the ReLU activation function in the hidden layer, effectively solving the problem of gradient disappearance and accelerating the model convergence speed. The calculation in the hidden layer linearly transforms the output of the previous layer through a formula and then processes it through the activation function, continuously extracting high-level features and enhancing the model's feature learning ability for data.

[0037] 3. Innovation in Intelligent Strategy Generation

[0038] Algorithm Optimization and Innovation: An improved Q - learning algorithm is adopted to generate intelligent strategies. An exploration factor based on confidence is introduced and combined with a genetic algorithm for optimization. The exploration factor increases the exploration of unknown actions in the initial stage of training to avoid falling into local optima. The genetic algorithm performs global search optimization on the strategies obtained by Q - learning to improve the strategy quality. For example, in the product recommendation strategy, it can dynamically adjust according to the real - time status of customers and prediction results, improving the pertinence and effectiveness of recommendations.

[0039] Policy Iterative Optimization and Innovation: Through continuous iterative learning, the strategy that can best improve the enterprise's goals is selected according to customer characteristics and predicted behaviors. During iteration, the learning rate, discount factor, and exploration factor coefficient are dynamically adjusted to balance the exploration and exploitation of the strategy. At the same time, combined with the genetic algorithm, simulating natural selection and genetic variation, selection, crossover, and mutation operations are performed on the strategy population to find the optimal solution in a broader strategy space, enhancing the adaptability and effectiveness of the strategy. Brief Description of the Drawings

[0040] Figure 1 It is a flowchart of a method for customer behavior prediction and intelligent strategy generation based on multi - source data fusion according to the present invention.

[0041] Figure 2 It is an architecture diagram of a system for customer behavior prediction and intelligent strategy generation based on multi - source data fusion according to the present invention. Detailed Embodiments

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Embodiment 1

[0043] Background of the Implementation Case: Taking a comprehensive financial service company as an example, the company has an online financial trading platform, offline business outlets, and social media official accounts. The online platform records customers' investment transaction records, wealth management product browsing history, account information, etc.; the offline business outlets store customers' face - to - face consultation records, account - opening materials, and offline business handling situations; the social media platform collects customers' evaluations of financial products, attention to industry dynamics, and interactions with the brand. The company hopes to integrate these multi - source data to predict customers' investment behaviors and formulate personalized financial service strategies. Data Collection and Pre - processing

[0044] 1. Data Collection

[0045] Extract the customer investment transaction records of the past three years from the PostgreSQL database of the online financial trading platform, including investment amount, investment time, investment product type, etc.; at the same time, obtain the browsing logs of customers' wealth management products, recording the product names browsed, browsing duration, etc.

[0046] Obtain the customer's account opening information from the SQL Server database of the offline business outlets, such as customer age, occupation, income level, etc.; and the offline business handling records, including financial questions consulted, business types handled, etc.

[0047] Use the API interfaces of social media platforms (such as Douyin, Xiaohongshu) to collect data such as video comments, picture - text likes, and sharing behaviors related to financial products posted by customers.

[0048] 2. Data cleaning

[0049] For the investment amount data, by calculating the inter - quartile range, it is found that some investment amounts are extremely small (possibly test data) or extremely large (possibly input errors), and these outliers are corrected or deleted. For example, after calculation, it is found that an investment amount is 0.01 yuan, which is far lower than the normal investment range, and it is deleted after verification as test data.

[0050] For the browsing duration data, remove the records with a browsing duration of less than 3 seconds, as these may be misoperations or invalid browsings. For example, in the wealth management product browsing logs, delete the records with a browsing duration of 1 second and 2 seconds.

[0051] Perform min - max normalization on numerical data such as the customer's investment amount and browsing duration, so that their value ranges are between [0, 1], which is convenient for subsequent model training. For example, after processing the investment amount data through the normalization formula, different magnitudes of investment amounts are unified into the [0, 1] interval, facilitating model learning. Data fusion

[0052] 1. Feature extraction

[0053] Extract the customer's investment frequency from the online transaction data 、average investment amount 、investment product preference and other features. For example, obtain the investment frequency by counting the number of investment times of customers within a period of time, calculate the average investment amount, and analyze the investment product type distribution to obtain the investment product preference.

[0054] Extract the customer's risk tolerance from the offline business outlet data (determined according to the risk assessment questionnaire filled out by the customer), business handling frequency , Types of Consultation Questions and other features. For example, according to the scores of the risk assessment questionnaire, the customer risk tolerance is divided into three levels: low, medium, and high.

[0055] Extract the customer's sentiment tendency from social media data (judged as positive, negative, or neutral through text analysis), interaction heat (the sum of the number of likes, comments, and shares), etc. Use natural language processing technology to perform sentiment analysis on social media comments to determine the sentiment tendency.

[0056] 2. Feature Fusion and Weight Determination

[0057] Concatenate the extracted feature vectors to obtain an initial fused feature vector. For example, concatenate the feature vector of online transaction data , the feature vector of offline business outlet data and the feature vector of social media data into a comprehensive feature vector.

[0058] Through multiple cross-validations, determine that the weight of online transaction data is 0.4, the weight of offline business outlet data is 0.3, and the weight of social media data is 0.3 to obtain the weighted fused feature vector. During the cross-validation process, divide the dataset into a training set, a validation set, and a test set, continuously adjust the weight combination, and determine the optimal weight according to indicators such as the prediction accuracy rate and recall rate of the model on the validation set. Construction and Training of Customer Behavior Prediction Model

[0059] Model Construction: Construct a deep neural network with 4 hidden layers. The input layer receives the fused feature vector. The number of nodes in the hidden layers is 128, 64, 32, and 16 respectively. The output layer outputs the probability that the customer will purchase a specific investment product within the next month. The number of neurons in each hidden layer gradually decreases. This structure helps the model gradually extract the high-level features of the data, avoid overfitting, reduce the computational amount at the same time, and improve the model training efficiency.

[0060] 2. Model Training: The model is trained using the Adam optimizer with a learning rate of 0.001 and 200 iterations. The Adam optimizer combines the advantages of the Adagrad and Adadelta optimization algorithms, which can adaptively adjust the learning rate for each parameter, showing a relatively fast convergence rate and good stability during training. Using historical investment transaction data as the training set, with whether the customer purchases a specific investment product as the label, a binary cross-entropy loss function is used for model training. During training, monitor the accuracy and loss values of the model on the validation set. After each iteration, calculate the binary cross-entropy loss between the predicted results of the model on the validation set and the true labels. When the validation set loss value does not decrease for 5 consecutive iterations, stop training, assuming that the model has reached a good convergence state to avoid overfitting. At the same time, record the accuracy change curve of the model during training to analyze the training effect of the model. Intelligent Strategy Generation and Application

[0061] Strategy Generation: Define the state space as the customer's features and predicted purchase probability, and the action space as the strategies that the financial service company can adopt, such as recommending high-risk and high-return investment products, recommending stable financial products, sending exclusive investment offers, and pushing personalized financial knowledge popularization, etc. Adopt the Q-learning algorithm with a learning rate of 0.1 and a discount factor of 0.9. In the Q-learning algorithm, the agent (here refers to the strategy generation module) continuously interacts with the environment (customers and the market), selects an action according to the current state, and will receive a reward and the next state after executing the action. The algorithm updates the Q value of the current state-action pair based on the reward and the Q value of the next state. For example, if a customer purchases a high-risk and high-return investment product after the recommendation and the purchase amount is large, give a positive reward (such as +10); if the customer does not purchase and expresses concern about the product's risk, give a negative reward (such as -5). Through multiple iterations of learning, obtain the optimal strategies in different customer states. For young customers with a high predicted purchase probability and high risk tolerance, recommend high-risk and high-return investment products and provide certain investment offers, such as reducing handling fees; for elderly customers with low risk tolerance, recommend stable financial products and push relevant financial knowledge popularization articles.

[0062] Strategy application: Apply the generated intelligent strategy to online financial trading platforms and offline business outlets. On the online platform, recommended investment products and services are dynamically displayed based on the customer's real-time browsing behavior and personal information; at offline business outlets, staff provide personalized consultation and services to customers based on the strategy generated by the system. Over a period of time, compare the customer investment conversion rate and customer satisfaction before and after the application of the strategy. Through data analysis, it was found that after the application of the strategy, the customer investment conversion rate increased by 15%, and customer satisfaction increased from 70% to 80%. Specifically, before the application of the strategy, the monthly investment conversion rate was 10%, which increased to 11.5% after application; the customer satisfaction survey results showed that before the application of the strategy, 70 out of 100 customers expressed satisfaction, and after the application, the number of satisfied customers increased to 80. This shows that the method and system proposed in the present invention can significantly improve the customer operation effect of financial service companies, and bring actual economic benefits and customer reputation improvement to enterprises. Example 2

[0063] Implementation case background: A comprehensive e-commerce platform has a massive amount of user data, including users' shopping records, browsing behaviors, search histories, and users' discussions on platform products and brands on social media. However, since the data is scattered in different systems, it is difficult to effectively integrate and utilize, resulting in challenges in precision marketing, product recommendations, and user retention. The platform hopes to use the technology of the present invention to predict user purchasing behavior through multi-source data fusion, formulate intelligent marketing strategies, and improve user shopping experience and platform sales.

[0064] 1. Data collection and preprocessing

[0065] Data collection: Obtain user transaction data for the past year from the MySQL database of the e-commerce platform, including order amount, purchased product category, purchase time, etc.; collect the platform's log files and parse information such as user browsed pages, browsing time, search keywords, etc.; use the API interface of social media platforms (such as Xiaohongshu, Weibo, and WeChat) to collect content related to platform products posted by users, such as product reviews, recommendations, and order sharing data.

[0066] Data cleaning: For order amount data, outliers are identified and processed by calculating the interquartile range, and abnormal order records with excessively large or small amounts are removed. For example, if an order amount is found to be far beyond the normal price range of similar products, it is corrected after verification as a data entry error. For browsing time data, records of less than 5 seconds are removed to avoid invalid browsing data interfering with analysis.

[0067] Data normalization: Perform min-max normalization on numerical data such as order amount and browsing duration. For example, map the order amount to the interval [0, 1] to make data of different magnitudes comparable and facilitate subsequent model training.

[0068] 2. Data fusion

[0069] Feature extraction: Extract features such as user purchase frequency, average purchase amount, and purchase category preference from transaction data; obtain features such as user browsing category preference, popular browsing pages, and browsing depth (number of browsing pages) from browsing data; extract features such as user sentiment towards different products (through text sentiment analysis) and social influence (number of fans, interaction volume) from social media data.

[0070] Feature fusion and weight determination: Concatenate the above-extracted feature vectors to form an initial fused feature vector. Through multiple cross-validation experiments, calculate metrics such as the prediction accuracy and recall rate of the model under different weight combinations, and determine that the weight of transaction data is 0.4, the weight of browsing data is 0.35, and the weight of social media data is 0.25. For example, through experiments, it is found that under a certain weight combination, the prediction accuracy of the model for user purchase behavior reaches 80%, and the recall rate reaches 75%, with the best comprehensive indicators. At the same time, according to the changes in platform business and the dynamic characteristics of data, re-perform cross-validation to adjust the weights every quarter.

[0071] 3. Construction and training of the customer behavior prediction model

[0072] Model construction: Build an improved deep neural network with 3 hidden layers. The input layer receives the fused feature vector, and the number of nodes in the hidden layers is 256, 128, and 64 respectively. The output layer outputs the probability that the user will purchase a specific product within the next week. The ReLU activation function is used in the hidden layer, and residual connections and attention mechanisms are introduced to enhance the model's learning ability for complex data features.

[0073] Model training: Use the Adagrad optimizer to train the model, set the learning rate to 0.01, and the number of iterations to 150. Use historical user data as the training set, take whether the user purchases a specific product as the label, and use the binary cross-entropy loss function for training. During the training process, monitor the accuracy and loss value of the model on the validation set, and stop training when the validation set loss value no longer decreases for 3 consecutive iterations to avoid overfitting. Record the accuracy change curve during the model training process to evaluate the model training effect.

[0074] 4. Intelligent strategy generation and application

[0075] Policy Generation: Define the state space as the user's characteristics and predicted purchase probability, and the action space as the strategies that the e-commerce platform can adopt, such as recommending relevant products, issuing coupons, pushing personalized advertisements, etc. Adopt an improved Q-learning algorithm, set the learning rate to 0.05, the discount factor to 0.85, and the exploration factor coefficient to 0.005. Give rewards according to the feedback of the user's purchase behavior. If the user purchases after the product is recommended, give a positive reward; if the user does not purchase and browses other competing products, give a negative reward. Through multiple iterations of learning, generate the optimal strategies under different user states. For example, for users with a high predicted purchase probability and who often purchase high-value products, recommend relevant products with high profits and issue large-value coupons; for new users who browse products but do not purchase, push exclusive discount advertisements for new users. At the same time, use the genetic algorithm to optimize the strategies generated by Q-learning. Set the population size of the genetic algorithm to 80, the crossover probability to 0.7, and the mutation probability to 0.15. After 15 generations of evolution, improve the quality of the strategies.

[0076] Policy Application: Apply the generated intelligent strategies to the e-commerce platform. When the user browses the product page, recommend products in real time according to the strategy; when the user logs in to the platform or on specific festivals, push personalized advertisements and coupons. Through data monitoring for a period of time, compare the user purchase conversion rate, average order value, and user retention rate before and after applying the strategy. After applying the strategy, the user purchase conversion rate increased by 18%, the average order value increased by 12%, and the user retention rate increased from 70% to 78%. Specific data shows that before applying the strategy, the daily purchase conversion rate was 5%, and after application, it increased to 5.9%; before applying the strategy, the average order value was 200 yuan, and after application, it increased to 224 yuan; before applying the strategy, the monthly new user retention rate was 70%, and after application, the new user retention rate increased to 78%. This shows that the present invention can significantly improve the operation effect and economic benefit of the platform in the application of the e-commerce industry. Practical Value

[0077] Data Integration and Insight Comprehensiveness: The present invention breaks the barriers between data through multi-source data fusion and comprehensively depicts customer characteristics. Integrate online transaction data, offline business outlet data, and social media data, enabling enterprises to understand customers from multiple dimensions, making up for the deficiencies of traditional methods based on single-source data analysis, and providing a rich data basis for in-depth insight into customer behavior.

[0078] Improved prediction accuracy: The deep neural network prediction model constructed based on the fused data can learn the complex patterns and features in the data. Through the automatic feature extraction and non - linear transformation of multiple hidden layers, the prediction accuracy of the model for customer behavior has been significantly improved. Compared with traditional prediction models, on the same test data set, the accuracy of the model of the present invention has increased by 10% - 20%, providing a reliable basis for enterprises to formulate precise strategies.

[0079] Enhanced strategy pertinence and effectiveness: The intelligent strategies generated based on reinforcement learning can be dynamically adjusted according to the real - time status of customers and prediction results. By continuously interacting and learning with customers and the market environment, the strategies can better meet the needs of different customers, improving the pertinence and effectiveness of the strategies. In practical applications, it has significantly improved the customer investment conversion rate and customer satisfaction of enterprises, enhancing the market competitiveness of enterprises.

[0080] Scalability and adaptability: The methods and systems of the present invention have good scalability and adaptability. In terms of data sources, it can easily access new data sources, such as third - party credit data, industry macro - data, etc., further enriching the data dimensions; in terms of models and algorithms, it can flexibly adjust the model structure and algorithm parameters according to the business needs and technological development of enterprises, adapting to different application scenarios and business objectives.

Claims

1. A customer behavior prediction and intelligent strategy generation method based on multi-source data fusion, characterized in that: The following steps are involved: Step 1: Collect customer data from multiple data sources; Step 2: Clean and normalize the collected data; Step 3: Perform feature fusion on the preprocessed data and determine the weight of each data source; Step 4: Construct a deep neural network prediction model based on the fused feature vector; Step 5: Train the prediction model and select appropriate activation function and loss function according to the prediction target; Step 6: Based on the prediction results, use reinforcement learning algorithm to generate intelligent strategies.

2. The method according to claim 1, characterized in that: In the data cleaning step, the interquartile range is used to identify outliers. The formula is: , the outlier judgment range is or .

3. The method according to claim 1, characterized in that: The data normalization adopts the minimum-maximum normalization method, and the formula is .

4. The method according to claim 1, characterized in that The multi-source data fusion adopts the method of feature splicing and weighted fusion. The fused feature vector ,in is the weight, For The feature vector extracted from the data source.

5. The method according to claim 1, characterized in that: The hidden layer in the deep neural network prediction model adopts the ReLU activation function, and the formula is .

6. The method according to claim 1, characterized in that If the probability of customer purchase is predicted, the output layer uses the sigmoid activation function, and the formula is , the loss function uses the binary cross entropy loss function .

7. The method according to claim 1, characterized in that The intelligent strategy generation adopts Q-learning algorithm, and the Q value update formula is: .

8. The method according to claim 1, characterized in that: The multiple data sources include relational databases, log files, and social media platform interfaces.

9. The method according to claim 1, characterized in that: During the data collection process, relevant data is extracted from different table structures of the relational database, and key information is parsed from log files.

10. The method according to claim 1, characterized in that When determining the weights of multi-source data, multiple cross-validation experiments are performed to calculate indicators such as the prediction accuracy of the model under different weight combinations, and the weight combination that optimizes the indicators is selected.

11. A customer behavior prediction and intelligent strategy generation system based on multi-source data fusion, characterized in that: include: Module 1: Data collection module, used to collect customer data from multiple data sources; Module 2: Data preprocessing module, used to clean and normalize the collected data; Module 3: Data fusion module, used to fuse the features of the preprocessed data and determine the weights; Module 4: Prediction model building module, used to build a deep neural network prediction model based on the fused feature vector; Module 5: Model training module, used to train the prediction model and select appropriate activation function and loss function; Module 6: Intelligent strategy generation module, used to generate intelligent strategies based on prediction results using reinforcement learning algorithms.

12. The system according to claim 11, characterized in that The data preprocessing module uses the interquartile range and minimum-maximum normalization methods to clean and normalize the data.

13. The system according to claim 11, characterized in that The data fusion module performs data fusion by means of feature splicing and weighted fusion.

14. The system according to claim 11, characterized in that The hidden layer of the deep neural network constructed by the prediction model building module adopts the ReLU activation function, and the output layer selects the activation function and the loss function according to the prediction target.

15. The system according to claim 11, characterized in that The intelligent strategy generation module uses Q-learning algorithm to generate intelligent strategies.

16. The system according to claim 11, characterized in that The data sources from which the data collection module collects data include relational databases, log files, and social media platform interfaces.

17. The system according to claim 11, characterized in that When collecting data, the data collection module extracts relevant data from different table structures of the relational database and parses key information from the log file.

18. The system according to claim 11, characterized in that When determining the weights, the data fusion module calculates indicators such as the prediction accuracy of the model under different weight combinations through multiple cross-validation experiments, and selects the weight combination that optimizes the indicators.

Images