Buried point analysis system oriented to customer behavior analysis

Abstract

The invention discloses a burying point analysis system for customer behavior analysis, and relates to the field of customer behavior analysis, the system comprises the following components: a data acquisition module, a digital twin model construction module, a business scene simulation module and a behavior prediction and decision support module; according to the invention, the digital twinborn model which accurately reflects customer behavior characteristics is constructed by using the digital twinborn model construction module and combining a deep learning algorithm and a data mining technology, and the model is imported into a diversified virtual business scene for behavior simulation through the business scene simulation module, so that the customer behavior simulation efficiency is improved. And finally, the behavior prediction and decision support module accurately predicts customer behaviors and generates a personalized decision plan by applying a space-time correlation prediction algorithm and an intelligent decision recommendation algorithm based on a knowledge graph, so that the customer behaviors can be accurately predicted. The functions jointly enhance the insight ability of enterprises for customer behaviors.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of customer behavior analysis, in particular to a point-in-time analysis system for customer behavior analysis. BACKGROUND

[0002] With the rapid development of Internet technology and the deepening of digital transformation, enterprises have increasingly strong needs for customer behavior insights, and customer behavior analysis has become a key means for enterprises to optimize product iteration, adjust marketing strategies, improve user experience, and enhance market competitiveness.

[0003] Traditional customer behavior analysis systems often face problems such as low data collection efficiency and quality, and insufficient system resource utilization when processing point-in-time data. On the one hand, due to the diversity and dynamics of customer behavior, fixed-interval data collection methods are difficult to adapt to different customer behavior patterns, leading to data redundancy or loss, affecting the accuracy of analysis results. On the other hand, traditional systems lack effective consideration of individual differences in customers during data processing and analysis, often using a one-size-fits-all analysis method, making it difficult to accurately capture the unique behavior characteristics and needs of each customer. In addition, traditional systems rely on simple statistical models or experience-based judgments when predicting customer behavior, lacking the ability to deeply mine and accurately predict complex customer behavior patterns, limiting the scientificity and rationality of enterprise decision-making.

[0004] In view of the problems of low data collection efficiency and quality, insufficient system resource utilization, and insufficient prediction accuracy of traditional customer behavior analysis systems, it is particularly important to develop a point-in-time analysis system for customer behavior analysis. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a point-in-time analysis system for customer behavior analysis, which can dynamically adjust the data sampling interval according to the customer behavior activity level by introducing an adaptive sampling interval algorithm, effectively improving the comprehensiveness and accuracy of data collection, and optimizing the efficiency of system resource utilization. In addition, the system combines deep learning algorithms and data mining techniques to build a digital twin model that accurately reflects customer behavior characteristics, and through business scenario simulation and behavior prediction and decision support modules, it realizes deep understanding and accurate prediction of customer behavior.

[0006] To solve the above technical problems, the present application provides the following technical solution: a point-in-time analysis system for customer behavior analysis, which comprises the following components: a data collection module, a digital twin model construction module, a business scenario simulation module, and a behavior prediction and decision support module. The data acquisition module collects various customer behavior data in real time by setting up tracking points on the product interface and application client, including but not limited to click behavior, browsing path, operation duration, purchase records, and search keywords. At the same time, it collects basic customer information to provide a data foundation for subsequent modeling. The digital twin model construction module: Based on the acquired customer historical behavior data and basic information, it uses deep learning algorithms and data mining technology to construct a digital twin model of each customer. Combined with the customer's basic information, through feature fusion algorithms, the model can comprehensively consider the individual differences of customers, thereby constructing a digital twin model that accurately reflects the customer's behavioral characteristics. The business scenario simulation module creates diverse virtual business scenarios based on the actual needs of enterprises. It imports the constructed digital twin models of individual customers into each virtual business scenario and simulates the behavioral decision-making paths of customers in different scenarios through a model-driven approach. The behavior prediction and decision support module analyzes and processes the behavior simulation results of the individual customer digital twin models in the business scenario simulation module. Using statistical methods and machine learning algorithms, it predicts how customers will respond to product iteration and marketing strategy adjustment decisions in actual business scenarios. The prediction results are presented to enterprise decision-makers in a visual form, and detailed decision-making plans are generated, including product improvement suggestions, marketing strategy optimization plans, and resource allocation plans, to assist enterprises in making scientific and reasonable decisions.

[0007] Furthermore, in the data acquisition module, when collecting customer behavior data, an adaptive sampling interval algorithm is used to collect the data. This algorithm dynamically adjusts the data sampling interval according to the activity level of customer behavior, thereby ensuring the accuracy of data collection while optimizing the use of system resources. Specifically, customer behavior activity is comprehensively evaluated by factors such as the number of behavioral events and the complexity of operations within a unit of time. The algorithm calculates the average value of customer behavior activity within a fixed window by statistically analyzing customer behavior data within that window, and uses this as the basis for adjusting the sampling interval. When a customer triggers a large number of behavioral events within a unit of time and the operations are highly complex, it indicates that the customer behavior is active. At this time, the algorithm will automatically shorten the sampling interval and collect data more frequently to ensure that every detail of the customer behavior can be captured. Conversely, when the number of customer behavioral events is small and the operations are simple, that is, when the behavior is in a flat state, the algorithm will increase the sampling interval to reduce unnecessary data collection, reduce data redundancy, and reduce the pressure on the system to store and process data. The window size for statistical behavior data is an optimal value selected through a large number of experiments, comparison of data integrity and calculation efficiency under different windows, and determination of the key parameter for adjusting the sampling interval in the algorithm according to the characteristics of the business scenario. The adaptive sampling interval algorithm can accurately adapt to the dynamic changes of customer behavior, improve the data collection efficiency and quality, and provide a better and more valuable data basis for subsequent digital twin model construction and other modules.

[0008] Further, in the digital twin model construction module, when extracting features from customer behavior data, a multi-scale attention feature extraction algorithm is used. First, the original behavior data is divided into different scale sub-sequences according to time sequence. For an original data sequence of length , it is divided into different scale sub-sequences . , Then, for each sub-sequence, the attention weight is calculated, and the attention weight calculation formula is: Wherein , is a self-defined similarity function, is the length of the sub-sequence. Through this algorithm, the key features of customer behavior under different scales can be focused, and the feature vector reflecting the customer behavior pattern and decision logic can be effectively extracted. Compared with traditional feature extraction methods, customer behavior features can be more comprehensively and accurately captured, and the accuracy and effectiveness of the digital twin model can be improved.

[0009] Further, in the digital twin model construction module, when combining customer basic information and behavior data for feature fusion, a dynamic weight fusion algorithm is used. Let the customer basic information feature vector be , the behavior data feature vector be , and the fused feature vector be The calculation formula is: Wherein and are dynamic weight coefficients, and the determination of the weight coefficients is based on the correlation between customer behavior and basic information. The correlation is calculated by mutual information, i.e. , , are the probability distributions of the features in , , and is the joint probability distribution. According to the mutual information value, the adaptive fuzzy control algorithm is used to adjust and When the mutual information value is high, increase the weight related to the basic information. Conversely, it increases. This allows the integrated features to reflect both individual customer differences and behavioral characteristics, thereby improving the accuracy of digital twins in simulating customer behavior.

[0010] Furthermore, in the business scenario simulation module, when creating virtual business scenarios, a scenario generation algorithm based on generative adversarial networks is used, and the generator... Receive random noise vector and business requirement condition vector Generate virtual business scenarios Discriminator Used to determine if the input scenario is a real-world scenario sample. Or a generated virtual scene During training, a simplified adversarial training objective function is used: in Indicates samples of real-world scenarios Expectation calculation, Represents the random noise vector and business requirement condition vector The expected value calculation of the combination is achieved by taking the logarithm of the discrimination probabilities of the real scene and the generated scene to construct the adversarial optimization objective of the generator and the discriminator, so as to prompt the generator to generate more realistic virtual business scenarios and the discriminator to improve its discrimination ability. To balance the training process of the generator and discriminator, an improved weight adjustment factor is introduced. The parameter update step size of the generator and discriminator is dynamically adjusted using the following formula: in The accuracy of the discriminator on the current batch of data, The accuracy at which the scene generated by the generator is identified as a real scene; when the discriminator accuracy is high, Increasing the value allows for a larger step size in the generator parameter updates, accelerating the generator's optimization speed; conversely, decreasing the value indicates that the generator is performing well. Decreasing the value gives the discriminator more opportunities to optimize, thereby achieving a dynamic balance in the training of both.

[0011] Further, in the business scenario simulation module, when the digital twin model is imported into the virtual business scenario for behavior simulation, a reinforcement learning driven model interaction algorithm is adopted, the state space S is defined as the current state set of the digital twin model in the scenario, including the behavior characteristics of the customer, the scenario parameters, the action space A is the set of behavior actions that the model can take, the reward function R(s, a) is used to evaluate the return after the model performs the action under the state , the model interacts with the virtual scenario, selects an action according to the state , obtains a reward and moves to a new state , the goal is to maximize the long-term cumulative reward: where is the discount factor, the value range is , the optimal value is searched within a certain range through the simulated annealing algorithm, the model parameter update adopts an improved algorithm based on the trust region strategy optimization, by limiting the amplitude of each policy update, the stability and convergence of the algorithm are guaranteed, so that the digital twin model can learn more reasonable behavior decision strategies in the virtual business scenario, and the authenticity and accuracy of the behavior simulation are improved.

[0012] Further, in the behavior prediction and decision support module, when the simulation results are analyzed and processed to predict customer behavior, a spatio-temporal correlation prediction algorithm is adopted, considering the correlation characteristics of customer behavior in time and space dimensions, for the time dimension, a time series model is constructed, an improved Transformer architecture is adopted, and a position encoding mechanism is introduced, the position encoding formula is: where is the position in the time series, is the dimension index, is the model dimension, for the spatial dimension, combined with the geographical location information of the customer, the influence weight of different spatial positions on the customer behavior is calculated through the spatial attention mechanism, the weight calculation formula is: where , are learnable feature mapping vectors, is the number of spatial positions, the information of time and space dimensions is fused, and the multi-layer perception is used for prediction to obtain the behavior probability distribution of the customer at different times and spaces in the future, so as to realize more accurate customer behavior prediction and provide more comprehensive basis for enterprise decision-making.

[0013] Furthermore, in the behavior prediction and decision support module, when generating decision-making plans, a knowledge graph-based intelligent decision recommendation algorithm is adopted. First, a knowledge graph related to customer behavior and business decisions is constructed. Nodes in the graph include customer behavior characteristics, business strategies, and market environment factors, while edges represent the relationships between nodes. For the predicted customer behavior results, retrieval and reasoning are performed in the knowledge graph. A graph neural network is then used to calculate the matching degree between different business decision-making plans and the current customer behavior and market environment. The matching degree calculation formula is as follows: in Indicates the first The decision-making options and the current situation are in the [number]th [year]. Matching degree in each dimension For activation function, For nodes The set of connected neighbor nodes, This is the weight matrix. The feature vectors of the neighboring nodes, Using a bias vector, decision-making schemes are sorted and filtered based on matching degree. Combined with the company's business objectives and resource constraints, personalized decision-making plans are generated. Explanations of the decision-making schemes are also provided to help corporate decision-makers understand the basis for their decisions and improve the acceptability and effectiveness of their decisions.

[0014] Furthermore, the data acquisition module, digital twin model construction module, business scenario simulation module, and behavior prediction and decision support module interact with each other through a communication mechanism improved based on a message queue telemetry transmission protocol. A data priority labeling and adaptive flow control strategy are introduced. Different priorities are assigned to different types of data, such as collected customer behavior data, model training parameters, and simulation results, based on their importance and timeliness. The priority calculation formula is as follows: in For data priority, The importance of data is quantified based on its role and impact within the system. This is a quantifiable value for data timeliness, reflecting the urgency at which the data needs to be processed. , The weighting coefficients are determined using the analytic hierarchy process. During communication, the data transmission rate is dynamically adjusted based on network bandwidth and system load. A fuzzy logic-based flow control algorithm is used to adjust the transmission window size according to network latency and packet loss rate indicators, ensuring the efficiency, stability, and real-time performance of data transmission between modules, and guaranteeing the overall normal operation and accuracy of the system analysis.

[0015] Compared with existing technologies, this data tracking analysis system for customer behavior analysis has the following advantages: I. This system utilizes a digital twin model building module, combined with deep learning algorithms and data mining techniques, to construct a digital twin model that accurately reflects customer behavioral characteristics. Through a business scenario simulation module, the model is imported into diverse virtual business scenarios for behavioral simulation. Furthermore, a reinforcement learning-driven model interaction algorithm enables the model to learn more reasonable behavioral decision-making strategies. Finally, the behavior prediction and decision support module employs spatiotemporal correlation prediction algorithms and knowledge graph-based intelligent decision recommendation algorithms to accurately predict customer behavior and generate personalized decision-making plans. These functions collectively enhance enterprises' ability to understand customer behavior, assisting them in making more scientific and reasonable decisions and improving market competitiveness.

[0016] Second, this system introduces an adaptive sampling interval algorithm, which dynamically adjusts the data sampling interval in the data acquisition module based on the activity level of customer behavior. When customer behavior is active, the system automatically shortens the sampling interval to capture more details, and when behavior is calm, the sampling interval is increased to reduce redundant data. This mechanism not only ensures the comprehensiveness and accuracy of data acquisition, but also significantly reduces the pressure on the system to store and process data, thereby optimizing the efficiency of system resource utilization.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is an overall flowchart of a data tracking analysis system for customer behavior analysis. Figure 2 This is a flowchart framework for key modules of a data tracking analysis system for customer behavior analysis. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1 Before a quarterly promotional event, a leading e-commerce platform discovered a significant gap in the user conversion path from browsing to adding to cart to placing an order. Especially during the peak promotional period when traffic surged, traditional marketing strategies struggled to accurately reach different types of users. The platform aimed to use a data tracking system to deeply analyze user behavior, optimize product recommendation logic and promotional activity design, and improve the overall conversion rate during the peak promotional period.

[0022] We deploy tracking points on core interfaces such as the platform homepage, product detail pages, and shopping cart checkout pages to collect real-time user behavior data, including the frequency of user clicks on product categories, the duration of browsing product details, records of adding and deleting products in the shopping cart, and search keywords. We also collect basic information filled in by users during registration, such as female users aged 25-35 who are more interested in beauty and personal care products, and users in North China who have higher requirements for the delivery time of fresh and cold chain products.

[0023] An adaptive sampling interval algorithm is used to optimize data collection efficiency: when the algorithm detects that a user has browsed 20 products continuously within 10 minutes and compared prices multiple times, it determines that the user is in a highly active state and automatically shortens the sampling interval to 5 seconds / time to ensure that every action of adding to cart or favorites is captured. If the user only opens the APP to browse the homepage and then remains idle for 30 minutes, the algorithm extends the sampling interval to 1 minute / time to reduce the storage of invalid data.

[0024] Based on user behavior data and basic information from the past six months, a multi-scale attention feature extraction algorithm is used to process the raw data. For example, the user's browsing history for the past 7 days is divided into different subsequences according to "daily time period" (morning, noon, and evening) and "product category" (clothing / digital / home furnishings). The algorithm automatically focuses on the user's behavior characteristics of frequently browsing digital products at 8 pm on weekdays, assigning higher attention weight to the data in this time period. The formula for calculating attention weight is: in , For a custom similarity function, is the length of the subsequence.

[0025] Combining basic user information and behavioral characteristics, an individual model is generated using a dynamic weight fusion algorithm. Let the feature vector of basic customer information be... The behavioral data feature vector is The fused feature vector The calculation formula is: wherein and are dynamic weight coefficients, adjusted by adaptive fuzzy control algorithm according to mutual information value and For a 30-year-old white-collar user with monthly consumption of 5000 yuan, the basic information characteristics of "high consumption ability" and the behavior characteristics of "high-frequency purchase of light luxury accessories" will adjust the weight through the adaptive fuzzy control algorithm, and the digital twin model finally generated will be more inclined to recommend brand goods with a single price of 800-1500 yuan.

[0026] A virtual promotion scene generation algorithm based on a generative adversarial network is adopted: generator receives a random noise vector and a business demand condition vector to generate a virtual business scene The discriminator is used to determine whether the input scene is a real scene sample or a generated virtual scene During the training process, a simplified adversarial training objective function is used: wherein represents the expected calculation of the real scene sample , represents the expected calculation of the combination of the random noise vector and the business demand condition vector ; To balance the training process of the generator and the discriminator, an improved weight adjustment factor is introduced, and the parameter update step of the generator and the discriminator is dynamically adjusted by the following formula: wherein is the accuracy of the discriminator on the current batch of data, is the accuracy of the scene generated by the generator being judged as a real scene, when the accuracy of the discriminator is high, the value increases, making the parameter update step of the generator larger, and accelerating the optimization speed of the generator, on the contrary, when the generator performs well, The value is reduced, giving the discriminator more opportunities to optimize. Input business requirements such as "full 300 minus 50", "limited time kill", "member exclusive discount", etc. The generator simulates page layout and traffic distribution scenarios under different discount intensity. For example, when generating a virtual scenario of "home focus group showing a makeup set of kill activities", the discriminator compares real traffic conversion data during the historical big promotion period to ensure the authenticity of the virtual scenario. During training, if the discriminator accurately identifies the virtual scenario for 3 consecutive times, the system will increase the generator's parameter update step size by improving the weight adjustment factor to accelerate the optimization of the realism of the virtual scenario.

[0027] After importing the user digital twin model into the virtual scene, the behavior is simulated through the model interaction algorithm driven by reinforcement learning: assuming that the model is in the state of "browsing the makeup kill activity page", the action space includes "clicking immediately purchase", "adding to shopping cart", "viewing user evaluation", etc. The reward function is set according to historical data - the conversion rate of users who click on the evaluation and then place an order is higher, so executing the "view evaluation" action will get a higher reward value. The model simulates the decision path of users under different promotion strategies by constantly interacting, such as 60% of users will choose to place an order after viewing more than 3 positive reviews.

[0028] Use the space-time correlation prediction algorithm to analyze the simulation results: consider the correlation characteristics of customer behavior in the time and space dimensions. For the time dimension, construct a time series model using an improved Transformer architecture and introduce a position encoding mechanism. The position encoding formula is: where is the position in the time series, is the dimension index, is the model dimension. For the spatial dimension, combine the customer's geographic location information and calculate the influence weight of different spatial positions on customer behavior through a spatial attention mechanism. The weight calculation formula is: where , are learnable feature mapping vectors, To determine the spatial location quantity, information from both temporal and spatial dimensions is fused and predicted using a multilayer perceptron to obtain the probability distribution of customer behavior in different times and spaces in the future. In the temporal dimension, based on the improved Transformer architecture, it is identified that the peak ordering period during major promotions is concentrated between 8 PM and 10 PM, and the location encoding mechanism strengthens the feature weights of time points such as "8 PM" and "9 PM". In the spatial dimension, combined with the user's delivery address, the spatial attention mechanism calculates that users in East China pay 1.8 times more attention to the "next-day delivery" service than those in Northwest China. After fusing spatiotemporal information, the model predicts that "pushing beauty flash sale products with the next-day delivery label at 8 PM in East China" will increase the click-through rate by 25%.

[0029] A knowledge graph-based intelligent decision-making and recommendation algorithm generation scheme is proposed as follows: First, a knowledge graph related to customer behavior and business decisions is constructed. Nodes in the graph include customer behavior characteristics, business strategies, and market environment factors, while edges represent the relationships between nodes. For predicted customer behavior results, retrieval and reasoning are performed within the knowledge graph. A graph neural network is then used to calculate the matching degree between different business decision options and the current customer behavior and market environment. The matching degree calculation formula is as follows: in Indicates the first The decision-making options and the current situation are in the [number]th [year]. Matching degree in each dimension For activation function, For nodes The set of connected neighbor nodes, This is the weight matrix. The feature vectors of the neighboring nodes, Using the bias vector, a knowledge graph is constructed containing nodes such as "user browsing time > 5 minutes", "previously purchased similar products", and "high price sensitivity during promotional periods". The edge relationship is defined as "positive correlation between browsing time and purchase intention". For the user group with "high intention but no order" in the prediction results, the algorithm recommends a combination strategy of "limited-time free shipping + free trial" after searching the graph. The graph neural network calculates that the matching degree between this strategy and user behavior reaches 82%. The final decision plan includes a specific product recommendation list (such as a certain brand of foundation), discount (30 yuan off) and push time (July 15, 20:00), and a visual explanation of the knowledge graph reasoning process to help the operations team understand "why this strategy is recommended".

[0030] Example 2 An internet finance platform discovered that middle-aged users aged 35-45 exhibited a phenomenon of "high pageviews but low conversion rates" when purchasing wealth management products. Furthermore, some users suddenly churned after browsing high-risk products. The platform hopes to use a data tracking system to uncover the investment preferences of this group, optimize product recommendation logic, reduce churn rate, and increase the asset allocation scale of high-net-worth users.

[0031] We deploy tracking points on pages such as "Homepage Recommendations," "Wealth Management Product Details," and "Risk Assessment" in financial apps to collect behavioral data such as the number of times users click on products with different risk levels, the duration of viewing historical return curves on product detail pages, operation records of adjusting investment periods, and search keywords. At the same time, we collect basic information such as users' occupation, annual income, and risk assessment results.

[0032] The adaptive sampling interval algorithm dynamically adjusts in this scenario: when a user continuously compares the historical maximum drawdown data of 3 R3-level wealth management products and saves the product details, the algorithm determines that the user is in a critical period of investment decision-making, and the sampling interval is shortened to 10 seconds / time to capture every data filtering action. If the user only opens the APP to check the account balance and then exits, the sampling interval is extended to 2 minutes / time to reduce data redundancy of low-frequency operations.

[0033] The algorithm uses a multi-scale attention feature extraction algorithm to process behavioral data: the user's investment records over the past three months are divided into subsequences based on "weekly transaction frequency" and "product type preference". The algorithm will focus on the behavior pattern of this middle-aged group redeeming financial products at the end of the quarter, and highlight the feature of "funds returning at the end of the quarter" through attention weight.

[0034] The dynamic weight fusion algorithm combines basic user information and behavioral characteristics: For a 40-year-old corporate executive with an annual income of 800,000, the basic information characteristic of "high risk tolerance" and the behavioral characteristic of "70% of funds allocated to equity funds in the past 6 months" will be adjusted by an adaptive fuzzy control algorithm. The resulting digital twin model is more inclined to recommend a mixed asset allocation scheme rather than a single low-risk product.

[0035] A scenario generation algorithm based on generative adversarial networks creates virtual investment scenarios: Inputting business requirements such as "equity-bond ratio 6:4", "linked to gold index", and "guaranteed minimum return of 2%", the generator simulates the return fluctuation curve scenarios of different asset portfolios. For example, when generating a virtual scenario of "mixed fund + gold ETF combination", the discriminator compares it with real historical market data to ensure the rationality of the return curve. When the scenario generated by the generator is misjudged by the discriminator as a real scenario, the system reduces the parameter update step size of the generator through an improved weight adjustment factor to avoid model overfitting.

[0036] The reinforcement learning driven model interaction algorithm simulates user decision-making: assuming that the model is in the state of "browsing the details page of a certain mixed fund", the action space includes "viewing portfolio details", "calculating investment yield", "consulting investment consultants", etc., and the reward function is set according to historical data - the probability of users buying after viewing portfolio details increases by 40%, so executing this action will obtain a higher reward, and the model simulates that 62% of middle-aged users will further consult investment consultants after viewing the top 10 holdings and the past performance of the fund manager.

[0037] The spatio-temporal correlation prediction algorithm analyzes the simulation results: in the time dimension, the improved Transformer architecture identifies that middle-aged users have stronger investment willingness at the beginning of the quarter, and the position encoding mechanism strengthens the time feature of "the first week of the quarter"; in the spatial dimension, combined with the high living cost pressure in the city where the user is located (such as a first-tier city), the spatial attention mechanism calculates that this group is more sensitive to the combination of "base product with guaranteed return + small proportion of high-risk product to seek yield", and the fused model predicts that "pushing the combination scheme of '60% stable wealth management + 30% stock fund + 10% gold ETF' to middle-aged users in first-tier cities at the beginning of April" will increase the purchase conversion rate by 35%.

[0038] The intelligent decision recommendation algorithm based on knowledge graph generates a scheme: a knowledge graph containing nodes such as "risk assessment result as balanced type", "redeemed stock fund in the past half year", "concerned about macroeconomic news", etc. is constructed, and the edge relationship is defined as "the correlation between redemption behavior and market fluctuations". For users who do not invest after redemption, the algorithm retrieves the graph and recommends a combination strategy of "fixed income + fund + market trend interpretation report", the graph neural network calculates a matching degree of 78%, the decision plan includes specific products, configuration proportion and touch mode, and through the knowledge graph visualization, it shows "why this combination can balance yield and risk", helping financial consultants accurately convey the value of the scheme to users.

[0039] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any equivalent embodiments with equivalent changes and modifications made to the above embodiments according to the technical essence of the present application, but without departing from the technical solution of the present application, are still within the scope of the technical solution of the present application.

Claims

1. A client behavior analysis-oriented in-app analytics system, characterized by, The system comprises the following components: a data acquisition module, a digital twin model construction module, a business scenario simulation module, and a behavior prediction and decision support module. The data acquisition module acquires various types of behavior data of customers in real time by setting a buried point on a product interface and an application client, including but not limited to click behavior, browsing path, operation duration, purchase record, search keyword, and simultaneously collects basic information of customers. The digital twin model construction module constructs an individual digital twin model of a customer based on the acquired historical behavior data and basic information of the customer, adopts a deep learning algorithm and a data mining technique, and combines the basic information of the customer to make the model consider individual differences of the customer through a feature fusion algorithm. The business scenario simulation module creates diversified virtual business scenarios according to actual needs of an enterprise, imports the constructed individual digital twin model of the customer into each virtual business scenario, and simulates a behavior decision path of the customer in different scenarios through model driving. The behavior prediction and decision support module analyzes and processes a behavior simulation result of the individual digital twin model of the customer in the business scenario simulation module, predicts a response of the customer to product iteration and marketing strategy adjustment in an actual business scenario by using a statistical method and a machine learning algorithm, presents the prediction result to an enterprise decision maker in a visualized form, and generates a detailed decision plan, including product improvement suggestion, marketing strategy optimization scheme, and resource allocation plan, to assist the enterprise in making a scientific and reasonable decision.

2. The system of claim 1, wherein, In the data acquisition module, the adaptive sampling interval algorithm is adopted to collect data when collecting behavior data of customers, and the algorithm dynamically adjusts a sampling interval of data according to a customer behavior activity level. Specifically, the customer behavior activity level is comprehensively evaluated by a number of behavior events in a unit time and an operation complexity factor, the algorithm calculates an average value of the customer behavior activity level in a fixed window by counting behavior data of the customer in the window, when a number of behavior events triggered by the customer in a unit time is large and the operation complexity is high, it is indicated that the customer behavior is in an active state, at this time, the algorithm automatically shortens the sampling interval to collect data more frequently, on the contrary, when the number of behavior events of the customer is small and the operation is simple, that is, the behavior is in a flat state, the algorithm increases the sampling interval to reduce unnecessary data collection, reduce data redundancy, and reduce the pressure of the system on data storage and processing. The window size for counting behavior data is an optimal value selected through a large number of experiments by comparing data integrity and calculation efficiency under different windows, and the key parameter for adjusting the sampling interval in the algorithm is determined according to characteristics of a business scenario. 3.The client behavior analysis oriented in-app analytics system of claim 1, wherein, In the digital twin model construction module, when extracting features from customer behavior data, a multi-scale attention feature extraction algorithm is used. First, the original behavior data is divided into subsequences of different scales according to the time series. For a length of... The original data sequence Divided into Subsequences of different scales , Then, for each subsequence, the attention weight is calculated using the following formula: wherein , is a self-defined similarity function, is a sub-sequence length.

4. The behavioral analytics client-oriented system of claim 1, wherein, In the digital twin model construction module, when the customer basic information and the behavior data are combined for feature fusion, a dynamic weight fusion algorithm is adopted, the customer basic information feature vector is , the behavior data feature vector is , and the fused feature vector is The calculation formula is: wherein and is a dynamic weight coefficient, adjusted by an adaptive fuzzy control algorithm according to the mutual information value and .

5. The behavioral analytics client-oriented system of claim 1, wherein, In the business scenario simulation module, when a virtual business scenario is created, a scenario generation algorithm based on a generative adversarial network is adopted Receiving a random noise vector And a business demand condition vector Generate a virtual business scenario Discriminator For judging whether the input scene is a real scene sample Or a generated virtual scene In the training process, a simplified adversarial training objective function is adopted: wherein represents the expected computation on real scene samples , represents the expected computation on random noise vectors and service demand condition vectors combined; To balance the training process of the generator and the discriminator, an improved weight adjustment factor is introduced The parameter update steps of the generator and the discriminator are dynamically adjusted by the following formula: wherein is the accuracy of the discriminator on the current batch of data, is the accuracy of the generator in generating scenes that are classified as real scenes by the discriminator, when the discriminator accuracy is high, is increased, which increases the step size of the generator parameter updates and speeds up the optimization of the generator, and conversely, is decreased, which gives the discriminator more opportunities to optimize.

6. The client behavior analysis oriented in-app analytics system of claim 1, wherein, In the business scenario simulation module, when the digital twin model is imported into the virtual business scenario for behavior simulation, a model interaction algorithm driven by reinforcement learning is adopted, the state space S is defined as the current state set of the digital twin model in the scenario, including the behavior characteristics of the customer, the scene parameters, the action space A is the set of behavior actions that the model can take, the reward function R(s, a) is used to evaluate the income after the model performs the action The next action is performed The model interacts with the virtual scenario, selects an action according to the state , obtains a reward and moves to a new state , and the goal is to maximize the long-term cumulative reward: wherein is a discount factor.

7. The behavioral analytics client-oriented system of claim 1, wherein, In the behavior prediction and decision support module, when analyzing and processing a simulation result to predict customer behavior, a spatio-temporal correlation prediction algorithm is adopted to consider the correlation characteristics of customer behavior in time and space dimensions, for the time dimension, a time series model is constructed, an improved Transformer architecture is adopted, and a position encoding mechanism is introduced, and a position encoding formula is: wherein is the position in the time series, is the dimension index, is the model dimension, for the spatial dimension, combined with the geographical location information of the customer, the influence weight of different spatial positions on the customer behavior is calculated through the spatial attention mechanism, and the weight calculation formula is: wherein 、 is a learnable feature mapping vector, is the number of spatial locations, the information of time and space dimensions is fused, and the behavior probability distribution of the customer at different times and spaces in the future is obtained through a multilayer perception machine.

8. The client behavior analysis oriented in-app analytics system of claim 1, wherein, In the behavior prediction and decision support module, when generating a decision plan, an intelligent decision recommendation algorithm based on a knowledge graph is adopted, a knowledge graph related to customer behavior and business decisions is first constructed, the nodes in the graph include customer behavior characteristics, business strategies, market environment factors, and the edges represent the association between the nodes, for the predicted customer behavior result, retrieval and reasoning are performed in the knowledge graph, the matching degree of different business decision schemes and the current customer behavior and market environment is calculated through a graph neural network, and the matching degree calculation formula is: in Indicates the first The decision-making options and the current situation are in the [number]th [year]. Matching degree in each dimension For activation function, For nodes The set of connected neighbor nodes, This is the weight matrix. The feature vectors of the neighboring nodes, Using a bias vector, decision-making schemes are ranked and filtered based on matching degree. Combined with the company's business objectives and resource constraints, personalized decision-making plans are generated, while explanations of the decision-making schemes are provided to help corporate decision-makers understand the basis for their decisions.

9. The client behavior analysis oriented in-app analytics system of claim 1, wherein, The data collection module, the digital twin model construction module, the business scenario simulation module and the behavior prediction and decision support module interact through a communication mechanism improved based on a message queue telemetry transport protocol, data priority marking and adaptive flow control strategies are introduced, different priorities are assigned to the collected customer behavior data, model training parameters and simulation results of different types of data according to their importance and timeliness, and the priority calculation formula is: wherein is a data priority, is a data importance quantization value, is a data timeliness quantization value, reflecting the urgency of data needing to be processed, , is a weight coefficient, in the communication process, the data sending rate is dynamically adjusted according to the network bandwidth and system load, a fuzzy logic-based flow control algorithm is adopted, and the sending window size is adjusted according to the network delay and packet loss rate indicators.

Images