AI-driven real-time customer behavior analysis response method and system

Abstract

The invention discloses an AI-driven real-time customer behavior analysis response method and system, and relates to the technical field of behavior analysis, and the method comprises the steps: collecting multi-channel user behavior data in real time, and constructing dynamic customer portrait information; performing multi-dimensional prediction based on the dynamic customer portrait information to generate a user behavior prediction result; driving the user behavior prediction result to automatically adjust a promotion strategy according to the AI, and generating personalized recommendation; the personalized recommendation is executed for real-time tracking, user feedback data is generated, updating response is conducted on the personalized recommendation according to the user feedback data, and a behavior analysis result is obtained. The technical problem that the promotion strategy matching degree is low due to inaccurate user behavior prediction in the prior art is solved, and the technical effect of improving the precision of the promotion strategy is achieved by collecting multi-channel user behavior data in real time, constructing a dynamic customer portrait, performing multi-dimensional prediction based on AI and automatically adjusting the promotion strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of behavior analysis, and particularly to an AI-driven real-time customer behavior analysis and response method and system. Background Art

[0002] In the current field of digital marketing and customer relationship management, enterprises usually rely on multi-channel data collection and analysis to optimize promotion strategies and improve user conversion rates. However, existing customer behavior analysis methods have many limitations. For example, traditional rule-based or simple statistical model methods are difficult to accurately predict user behavior, resulting in a low matching degree of promotion strategies and affecting marketing effects. In addition, due to the highly dynamic and personalized characteristics of user behavior, it is difficult to adjust promotion strategies in real time, resulting in lagging promotion content and unable to meet the immediate needs of users. Summary of the Invention

[0003] This application provides an AI-driven real-time customer behavior analysis and response method and system for solving the technical problems of limited applicability and low automation in existing technologies during document scanning.

[0004] In view of the above problems, this application provides an AI-driven real-time customer behavior analysis and response method and system.

[0005] In the first aspect of this application, an AI-driven real-time customer behavior analysis and response method is provided. The method includes:

[0006] Real-time collection of multi-channel user behavior data to construct dynamic customer portrait information; multi-dimensional prediction based on the dynamic customer portrait information to generate user behavior prediction results; automatic adjustment of promotion strategies according to the AI-driven user behavior prediction results to generate personalized recommendations; execution of the personalized recommendations for real-time tracking to generate user feedback data, and update response to the personalized recommendations according to the user feedback data to obtain behavior analysis results.

[0007] In the second aspect of this application, an AI-driven real-time customer behavior analysis and response system is provided. The system includes:

[0008] A portrait information construction module for real-time collection of multi-channel user behavior data to construct dynamic customer portrait information; a multi-dimensional prediction module for multi-dimensional prediction based on the dynamic customer portrait information to generate user behavior prediction results; a personalized recommendation generation module for automatically adjusting promotion strategies according to the AI-driven user behavior prediction results to generate personalized recommendations; an update response module for executing the personalized recommendations for real-time tracking to generate user feedback data, and update response to the personalized recommendations according to the user feedback data to obtain behavior analysis results.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] This application collects multi-channel user behavior data in real time to construct dynamic customer portrait information; makes multi-dimensional predictions based on the dynamic customer portrait information to generate user behavior prediction results; automatically adjusts the promotion strategy according to the AI-driven user behavior prediction results to generate personalized recommendations; executes the personalized recommendations for real-time tracking to generate user feedback data, and updates and responds to the personalized recommendations according to the user feedback data to obtain behavior analysis results. The present invention solves the technical problem in the prior art that the user behavior prediction is inaccurate, resulting in a low matching degree of the promotion strategy. By collecting multi-channel user behavior data in real time, constructing a dynamic customer portrait, making multi-dimensional predictions based on AI, and automatically adjusting the promotion strategy, the technical effect of improving the accuracy of the promotion strategy is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0012] Figure 1 It is a schematic flowchart of an AI-driven real-time customer behavior analysis and response method provided for the embodiments of this application;

[0013] Figure 2 It is a schematic structural diagram of an AI-driven real-time customer behavior analysis and response system provided for the embodiments of this application.

[0014] Explanation of reference numerals: Portrait information construction module 11, multi-dimensional prediction module 12, personalized recommendation generation module 13, update response module 14. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] By providing an AI-driven real-time customer behavior analysis and response method and system, this application aims to solve the technical problem in the prior art that the user behavior prediction is inaccurate, resulting in a low matching degree of the promotion strategy. By collecting multi-channel user behavior data in real time, constructing a dynamic customer portrait, making multi-dimensional predictions based on AI, and automatically adjusting the promotion strategy, the technical effect of improving the accuracy of the promotion strategy is achieved.

[0016] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts belong to the scope of protection of the present application.

[0017] It should be noted that any variations of the terms "including" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0018] Embodiment 1, as Figure 1 shown, the present application provides an AI-driven real-time customer behavior analysis and response method, and the method includes:

[0019] Step S100: Real-time collect multi-channel user behavior data and construct dynamic customer portrait information.

[0020] In the embodiments of the present application, dynamic customer portrait information is constructed by real-time collecting multi-channel user behavior data. Specifically, first, access the user behavior data source through the API interface to collect the original data and form an initial customer portrait. Subsequently, perform feature extraction on the original data to generate a structured behavior feature set, and construct a behavior feature time series based on time series analysis. Then, perform dynamic weight allocation on the feature data according to the time series, extract key attributes, and generate user attribute tags. Finally, use these attribute tags to continuously update the initial customer portrait, and finally form accurate and real-time updated dynamic customer portrait information.

[0021] Furthermore, in the method provided by the embodiments of the application, the real-time collection of multi-channel user behavior data and the construction of dynamic customer portrait information further include:

[0022] Real-time access the user behavior data source through the API interface for data collection to obtain the original data, where the original data includes initial customer portrait information; perform feature extraction on the original data to generate a structured behavior feature set, perform time series analysis based on the structured behavior feature set to generate a behavior feature time series; perform dynamic weight allocation on the structured behavior feature set according to the behavior feature time series, and generate user attribute tags according to the weight allocation result; update the initial customer portrait information based on the user attribute tags to construct the dynamic customer portrait information.

[0023] In the embodiments of the present application, first, the user behavior data sources are accessed in real time through the API interface, including website, mobile application, and third-party platform data, and data collection is performed. During the data collection process, the website side records the user's access path, stay time, and click behavior through the buried point technology. The mobile application stores the user's usage duration, interface switching, and interaction operations using the log recording method. The third-party platform data obtains the user's interaction information on social media or e-commerce platforms, such as likes, shares, comments, etc., through the open API. These data form the raw data after collection, which contains the user's basic information, browsing trajectory, interaction frequency, and historical transaction records, as the preliminary customer portrait information.

[0024] After obtaining the raw data, feature extraction is performed on it to convert the unstructured data into a standardized structured behavior feature set. This process uses the rule matching method and classifies different user behaviors through predefined data classification rules. For example, the record of accessing the product page of the e-commerce platform can be classified as "browsing behavior", likes or comments can be classified as "interaction behavior", and purchase records can be classified as "transaction behavior". At the same time, for numerical data, such as access frequency, purchase amount, etc., direct numerical normalization processing is performed to ensure that all data formats are unified and applicable to subsequent analysis. Finally, a structured behavior feature set is generated, which contains information such as the user's behavior category, behavior frequency, and occurrence time.

[0025] Based on the structured behavior feature set, time series analysis is performed to obtain the pattern of user behavior changes over time. This process uses the sliding window statistical method and accumulatively calculates the user behavior data by setting a fixed time window, such as 7 days or 30 days. For example, the number of user browsing times, interaction frequency, and purchase behavior are statistically analyzed within a 7-day window to form a behavior feature time series. This time series is used to analyze the short-term interest changes of users. For example, whether a certain user has increased the number of browsing times for a certain type of product in the past week, or whether the interaction frequency of specific content has decreased. By means of the sliding window, the user's behavior change trend is continuously updated to provide data support for subsequent analysis.

[0026] After obtaining the behavior feature time series, dynamic weight assignment is performed according to the time impact degree of the user behavior to ensure that the influence of recent behavior is greater. This process uses the exponential decay function method to assign different weights to the user's behavior at different time points. For example, the behavior weight of the most recent 1 day is the highest, the behavior weight of the past 7 days is the second highest, and the behavior weight 30 days ago is the lowest. The weight calculation formula is M t = M0·e -λt where M t$M_t$ represents the behavior weight at time $t$, $\lambda$ is the decay coefficient, and $M_0$ is the initial weight, which is usually set to 1 or assigned based on the starting importance of user behavior. For example, if a certain type of behavior (such as purchase behavior) is considered to have a greater impact on the user profile, the initial weight of this behavior can be set to a higher value, such as $M_0 = 2$; while for behaviors with lower impact (such as page views), it can be set to $M_0 = 1$. When setting the decay coefficient $\lambda$, consider the change rate of user behavior and the response speed of the recommendation system. For example, in the field of fast-moving consumer goods, such as food and clothing, users' interests change relatively quickly, so a larger decay coefficient can be set, such as $\lambda = 0.5$, so that the behavior weight a week ago decreases rapidly. In the field of durable consumer goods, such as cars and electronic products, users' purchase decision-making cycles are longer, so a smaller decay coefficient can be set, such as $\lambda = 0.1$, to retain the behavior impact within a longer time range. Through dynamic weight allocation, according to the temporal characteristics of user behavior, behaviors at different time points are weighted, and the weight distribution of each behavior feature is calculated, providing a basis for the subsequent generation of user attribute tags.

[0027] After that, based on the weight allocation results, user attribute tags are generated to further refine the user profile. This process uses a threshold screening method, and thresholds are set for classification according to the cumulative values of behavior data in different time windows. For example, if the number of times a user views a certain type of product within the last 7 days exceeds the set threshold, it is marked as "recently interested", and if the number of purchases exceeds a certain number, it is marked as "high-frequency purchaser". In addition, the interaction frequency of the user can also be combined to set tags such as "highly active user" and "lowly active user". These user attribute tags enable the system to accurately describe users' interests and habits based on their actual behavior characteristics.

[0028] Finally, based on the generated user attribute tags, the initial customer profile information is updated to construct dynamic customer profile information. This update process uses a weighted cumulative update method, and the original customer profile is adjusted by combining the newly generated user attribute tags. For example, if a user's browsing behavior in a specific category has increased recently, the interest weight of this category is increased; if the interaction frequency of a certain category has decreased, its weight is decreased. In addition, the activity score is also adjusted according to recent behaviors to ensure that the user profile can accurately reflect the current state of the user. Through the update process, the customer profile is continuously optimized to dynamically adapt to the changes in users' interests, and finally a complete dynamic customer profile information is formed, providing high-quality data support for subsequent personalized recommendations and precision marketing.

[0029] Step S200: Perform multi-dimensional prediction based on the dynamic customer profile information to generate user behavior prediction results.

[0030] In the embodiments of the present application, multi-dimensional prediction is performed based on dynamic customer profile information to generate user behavior prediction results, so as to improve the accuracy of personalized recommendation and the effectiveness of promotion strategies. Specifically, first, context analysis is performed based on the time series of behavioral characteristics to determine the associated data of user behavior, and these data are fused to construct a pre-trained intent classification model. Subsequently, the dynamic customer profile information of the user is input into the model to calculate the intent probability distribution information of the user, thereby predicting the possible behavior trends of the user. Then, a probability distribution threshold is set to screen the intent probabilities, and only the intents with high confidence are retained, and high-confidence intent labels are generated. Finally, these intent labels are used to perform identity deduction on the user profile, further predicting the potential needs of the user, and finally forming the user behavior prediction results to provide support for precision marketing and personalized recommendation.

[0031] Further, in the method provided by the application embodiments, when multi-dimensional prediction is performed based on the dynamic customer profile information to generate user behavior prediction results, it further includes:

[0032] Performing context analysis based on the time series of behavioral characteristics to determine context-associated data, fusing the context-associated data to construct a pre-trained intent classification model; synchronizing the dynamic customer profile information to the pre-trained intent classification model to obtain intent probability distribution information; setting a probability distribution threshold, and screening by traversing the intent probability distribution information according to the probability distribution threshold to generate high-confidence intent labels; performing identity deduction on the dynamic customer profile information according to the high-confidence intent labels to generate the user behavior prediction results.

[0033] In the embodiments of the present application, first, the sequence pattern mining method (SPADE algorithm) is used to analyze the time series of user behavioral characteristics to determine context-associated data. SPADE is a sequence pattern mining algorithm that can discover high-frequency behavioral patterns in time series data. For example, by analyzing the browsing, adding to cart, and placing an order behaviors of users, if it is found that the pattern of "browsing A - browsing B - adding A to cart - placing an order for A" appears repeatedly among multiple users, it can be inferred that there is a behavioral association between A and B, and A is more likely to be the final purchased product. Through the SPADE algorithm, context-associated information of user behavior is mined, such as "searching for mobile phones first and then searching for accessories", "browsing low-price products first and then browsing high-price products", etc. Through context analysis, context-associated data between user behaviors is obtained.

[0034] Next, based on the determined context-related data, the feature normalization method (Min-Max normalization) is adopted for data fusion, and a pre-trained intent classification model is constructed. Min-Max normalization is a linear transformation method that can map feature data with different dimensions to the same range (such as 0 to 1), eliminating the influence caused by different data scales. For example, normalizing the number of user browsing times (0 - 500 times), purchase amount (0 - 10,000 yuan), and social interaction times (0 - 100 times) so that their values are distributed between 0 and 1, ensuring the comparability of data during calculation. Based on the normalized data, a hierarchical classification method is used to construct the intent classification model, that is, according to different user behavior patterns, the intents are divided into multiple levels, such as "high purchase intent", "low purchase intent", "high churn risk", etc. Through this step, a pre-trained intent classification model is obtained, which can classify the behavior patterns of different users and provide support for subsequent probability calculations.

[0035] Subsequently, the rule matching method is adopted. The dynamic customer profile information is input into the pre-trained intent classification model, and the intent probability distribution information of the user is calculated. The intent probability distribution information includes purchase intent, product attention, and churn risk level, that is, different behaviors that the user may take and their corresponding probabilities. For example, if a user has browsed a certain brand of mobile phone more than 10 times in the last 7 days and has searched for related accessories of this brand, then according to the pre-trained rules, it can be matched to the "high purchase intent" category, and the purchase probability is calculated as 85%. The rule matching method is based on artificially set behavior patterns and weights, such as "number of browsing times in the last 7 days > 10, purchase intent probability = 70% + recent interaction weighted value", so as to calculate the possibility of a user's behavior. Through rule matching, the intent probability distribution information of each user, that is, the occurrence probabilities of different behaviors, is obtained, providing a basis for subsequent high-confidence intent screening.

[0036] After obtaining the intent probability distribution information, the threshold screening method is adopted to traverse and screen the intent probabilities, and only high-confidence intent labels are retained. Threshold screening is a commonly used classification method, that is, a fixed threshold is set, and data below this threshold is filtered. For example, if the threshold for purchase intent is set at 80%, then all users with a purchase intent probability higher than 80% will be labeled as "high-intent purchase users", while users with a probability lower than 80% will not be included in this round of prediction scope. Similarly, for the churn risk level, if a user's churn probability exceeds 90%, then this user will be labeled as a "high-risk churn user". Through threshold screening, high-confidence intent labels are obtained, that is, accurately identifying which users may purchase, pay attention to a product, or are about to churn, providing reliable data support for subsequent user behavior prediction.

[0037] Finally, based on the time series retrospective analysis method, combined with high-confidence intent tags, the dynamic customer portrait information is identified and deduced, and finally the user behavior prediction result is generated. Time series retrospective analysis is a method of inferring future trends based on time order, which can predict users' future behaviors according to their historical behavior patterns. For example, if a user's purchase intention has shown an increasing trend in the past 30 days (70% → 75% → 80% → 85%), it is speculated that the probability of the user completing a purchase in the next week will further increase. On the contrary, if a user's churn risk level has been continuously rising in the past 30 days, it is judged that the user may reduce the frequency of using the platform in the near future or even stop using it. Through the identification and deduction, the user behavior prediction result is obtained, providing a decision-making basis for precision marketing, personalized recommendation, and user recall strategies.

[0038] Step S300: Automatically adjust the promotion strategy according to the AI-driven user behavior prediction result to generate personalized recommendations.

[0039] In the embodiment of the present application, when automatically adjusting the promotion strategy according to the AI-driven user behavior prediction result, first, a preset strategy space is constructed, and a promotion strategy is randomly selected therein to provide multiple possible marketing plans. Subsequently, a sliding time window mechanism is used to perform feature engineering processing on the time series of behavioral characteristics to extract the time series behavioral feature vectors to capture the trend of users' behaviors changing over time. Then, based on the time series behavioral feature vectors, user interest decay analysis is performed on the dynamic customer portrait, and an incremental portrait update package is generated to ensure that the recommended content conforms to the user's latest interest preferences. After that, the incremental portrait update package is further classified to determine the user intent classification result, and the result is mapped to the preset strategy space for matching to generate a strategy matching result. Finally, the promotion strategy is automatically optimized according to the strategy matching result to generate personalized recommendations.

[0040] Furthermore, in the method provided by the embodiment of the application, automatically adjusting the promotion strategy according to the AI-driven user behavior prediction result to generate personalized recommendations further includes:

[0041] Construct a preset strategy space, make a random selection based on the preset strategy space to determine the promotion strategy; use a sliding time window mechanism to perform feature engineering processing on the time series of behavioral characteristics to extract the time series behavioral feature vectors; perform user interest decay analysis on the dynamic customer portrait according to the time series behavioral feature vectors to generate an incremental portrait update package; classify based on the incremental portrait update package to determine the user intent classification result, map the user intent classification result to the preset strategy space for matching to generate a strategy matching result; automatically optimize the promotion strategy according to the strategy matching result to generate the personalized recommendation.

[0042] In the embodiments of the present application, first, a policy pool construction method is adopted to construct a preset policy space, that is, all available promotion policies are collected to provide different marketing options. The preset policy space is a set containing various promotion methods, such as discount offers, product recommendations, targeted advertisement push, email marketing, SMS notifications, etc. Through the random sampling method, different promotion policies are selected from the policy space to ensure that different users can receive diversified marketing plans. For example, for new users, "first order discount" or "exclusive recommendation for new users" is randomly selected as the promotion policy, while for active users, "personalized product recommendation" or "points redemption reminder" is selected. By constructing the preset policy space, a set of promotion policies available for optimization is obtained, providing a basis for subsequent matching and adjustment.

[0043] Subsequently, the sliding time window mechanism is used to segment the original behavior data for analyzing user behavior characteristics in different time dimensions. The sliding time window mechanism is a time series analysis method that continuously updates and analyzes time series data by setting a fixed window length. The window length is dynamically adjusted in the range of 5 minutes to 24 hours according to the business scenario, and the sliding step is set to 1 / 10 of the window length to ensure a balance between the frequency of data update and computational efficiency. For example, in a short-term high-frequency interaction scenario (such as e-commerce limited-time flash sales), the window length can be set to 5 minutes to quickly respond to user behavior changes; while in a long-term consumption trend analysis scenario (such as high-value product purchase decisions), the window length can be set to 24 hours to capture changes in user purchase intentions. Through the sliding time window mechanism, the behavior data of users in different time dimensions is extracted to generate a time series behavior feature vector, providing support for subsequent user interest analysis.

[0044] After obtaining the time series behavior feature vector, further user interest decay analysis is carried out to calculate the changing trend of users' attention to different product categories or services. This process is based on user activity calculation and the double exponential decay model to model the change of users' interest over time, so as to avoid recommending outdated or less interesting content to users. First, the tanh function is used to calculate the user activity decay factor. Based on this, the double exponential decay function is further used to calculate the changing trend of users' interest to ensure that the recommendation can adapt to the dynamic changes of users' interests. Through this method, the historical feature weights in the dynamic customer portrait are updated, and different features are differentially compressed to generate an incremental portrait update package.

[0045] Subsequently, user intent classification is performed based on the incremental portrait update package, and the classification results are mapped to a preset policy space for matching to generate a policy matching result. This process uses a rule-based classification method, that is, users are classified according to their behavioral characteristics and mapped to appropriate promotion strategies. For example, based on the user's recent browsing and purchase behaviors, users are classified into categories such as "potential purchasers", "comparative decision-making users", "users about to churn", etc., and appropriate promotion strategies are matched for each category. For example, for "potential purchasers", "time-limited discounts" or "exclusive coupons" are recommended; for "comparative decision-making users", "comparative analysis of similar products" or "summary of user reviews" are recommended; for "users about to churn", "recall emails" or "personalized push reminders" are recommended. Through user intent classification, a policy matching result is obtained, enabling promotion strategies to be precisely matched according to the behavioral characteristics of different users.

[0046] Finally, based on the policy matching result, the promotion strategy is automatically optimized, and finally personalized recommendations are generated. This optimization process uses an adaptive decision-making engine, which adjusts the promotion strategy by continuously analyzing the user's feedback on the recommended content. For example, if a user has a high click-through rate on the personalized recommended products, the recommendation weight of this category of products is increased; if a user does not respond to the marketing information pushed multiple times, the push frequency is reduced or the recommended content is replaced to avoid user churn. The adaptive decision-making engine is based on the reinforcement learning method, enabling the promotion strategy to be continuously optimized. For example, the recommendation ranking is adjusted using the user's historical click data, so that the products or services that best match the user's interests are ranked at the top. Finally, the optimized personalized recommendations are output, including accurate product recommendations, personalized discount pushes, targeted advertisement displays, etc. Through the automatic optimization of the promotion strategy, highly accurate personalized recommendations are generated.

[0047] Furthermore, in the method provided by the application embodiment, when performing user interest decay analysis on the dynamic customer portrait according to the time-series behavior feature vector to generate an incremental portrait update package, it further includes:

[0048] Mapping is performed based on the time-series behavior feature vector to establish user activity, and a decay factor is obtained. The formula is:

[0049]

[0050] where α is the decay factor, A is the user activity determined according to the time-series behavior feature vector, and the time-series behavior feature vector is in a direct proportion relationship with the user activity. A avg is the historical average value of the user activity, and σ Ais the standard deviation; updating the historical feature weight values of the dynamic customer portrait according to the attenuation factor to generate multiple feature update weight values; performing differential compression according to the multiple feature update weight values to generate a feature change matrix, and aligning the feature change matrix according to the time stamp to generate the incremental portrait update package.

[0051] In the embodiments of the present application, first, a user activity mapping is established based on the time-series behavior feature vector to quantify the interaction intensity of the user within a specific time range to measure the interest weight of different features. Calculate the user activity A to measure the recent behavior frequency and interest intensity of the user. The calculation of the activity is based on the interaction data such as browsing, clicking, and purchasing of the user, and different weights are assigned to different behaviors. The calculation formula is A = W1·N view +W2·N click +W3·N purchase , where N view , N click , N purchase respectively represent the browsing, clicking, and purchasing times of the user; W1, W2, and W3 are the weights of different interaction behaviors, which are preset by technical experts. To smooth the influence of short-term behaviors, a 30-day moving average is used to calculate the historical activity, and the formula is where A represents the daily activity score of the user in the past 30 days, and this average value is used to determine whether the user's current activity is higher or lower than its historical average level, so as to determine whether the user's interest needs to be adjusted.

[0052] After that, calculate the attenuation factor α to control the update rate of the feature weight and ensure that the adjustment of the user's interest conforms to the behavior trend. where α is the attenuation factor, A is the user activity determined according to the time-series behavior feature vector, the time-series behavior feature vector is in a direct proportion relationship with the user activity, A avg is the historical average value of the user activity, σ A is the standard deviation. This formula uses the hyperbolic tangent (tanh) function for normalization to ensure that the value range of α is between (-1, 1); if A > A avg (that is, the user's recent activity has increased), then α takes a larger value, indicating that the interest attenuation speed should be reduced and more historical behavior information should be retained; if A < A avg (that is, the user's recent activity has decreased), then α takes a smaller value, indicating that the interest attenuation speed should be accelerated and less relevant recommended content should be reduced.

[0053] After obtaining the attenuation factor α, a weighted update method is used to update the historical feature weight values of the dynamic customer portrait, enabling the user portrait to retain long-term interests and quickly adapt to short-term changes; the formula is w t = α·w t-1+(1 - α)·Δx t ; where w t is the latest feature weight calculated at the current time step t, w t-1 is the feature weight of the previous time step, α is the decay factor, and Δx t is the decay rate. Δx t is calculated from the deviation between the user's recent interaction frequency with this feature and the historical mean. For example, the change trend is obtained by comparing the feature activity in the most recent 7 days with that in the most recent 30 days, and a threshold is set to control the change rate.

[0054] After completing the feature weight update, compression is performed according to the differences in the updated weight values of the features to generate a feature change matrix for extracting the key interest change trend. The feature change matrix is used to record the differences in the user interest features at different time steps and normalize the increase and decrease changes of the features, enabling the interest fluctuations of different users to be compared on a unified scale. For example, if a user's weight change for a certain feature in the most recent 7 days is large (such as increasing from 0.3 to 0.8), the growth trend of this feature is marked in the feature change matrix so that it can be preferentially processed in the recommendation system; conversely, if the weight of a certain feature continuously decreases over multiple time steps, the weight of this feature is reduced or removed to reduce ineffective recommendations.

[0055] Finally, the feature change matrix is aligned according to the time stamps to generate an incremental portrait update package, ensuring that the personalized recommendation system can reflect the latest state of the user's interest in real time. Time stamp alignment is used to normalize the feature change data at different time steps, enabling the system to optimize based on the latest behavior trends when updating the user portrait. For example, in scenarios with significant short-term trends (such as during promotional activities), the feature changes in the most recent 1 - 3 days are given priority, while in scenarios with long-term stable interests (such as long-term subscription services), adjustments are made based on the change trends in the most recent 30 days. The incremental portrait update package contains the user's latest interest weights, feature change trends, and recommendation strategy adjustment parameters, enabling the system to perform personalized optimization based on the latest user behavior in the next recommendation process.

[0056] Finally, the decay factor is calculated by establishing a mapping of user activity, and the historical feature weights are updated based on the decay factor. Subsequently, the differences in the changes of the feature weights are compressed to generate a feature change matrix, and the feature change matrix is aligned according to the time stamps. Finally, an incremental portrait update package is generated to ensure that the recommendation system can provide accurate personalized recommendations in an environment where the user's interests change dynamically, improve the recommendation effect, and optimize the user experience.

[0057] Step S400: Perform real-time tracking of the personalized recommendation to generate user feedback data, and update and respond to the personalized recommendation based on the user feedback data to obtain a behavior analysis result.

[0058] In the embodiments of the present application, by tracking user responses in real time, user behavior data is captured, and policy execution tracking is performed based on user response metrics to evaluate the actual effect of personalized recommendations. A policy effect evaluation report is generated according to the policy execution situation, and a policy effect feedback matrix is constructed through data mapping and added to the user feedback data as the core basis for subsequent optimization of the recommendation policy. Based on the user feedback data, the personalized recommendation is optimized using reinforcement learning methods. The model is trained through the policy effect feedback matrix to generate learning results, and then confidence analysis is performed to construct multiple confidence intervals to measure the reliability of the recommendation policy among different user groups. All confidence intervals are traversed, and a comparative analysis of the recommended content and the feedback data is performed to extract the recommended data set that needs to be updated and mark it as the data group to be updated. Finally, the influence coefficient is calculated using behavior impact analysis, and the personalized recommendation policy is adjusted according to the influence coefficient to ensure that the recommendation system can accurately adapt to changes in user interests, and finally obtain the behavior analysis result.

[0059] Further, in the method provided by the application embodiments, when performing the personalized recommendation for real-time tracking to generate user feedback data, it further includes:

[0060] When performing the personalized recommendation for capture to obtain user response metrics, policy execution tracking is performed according to the user response metrics to generate a policy effect evaluation report; data mapping is performed based on the policy effect evaluation report to construct a policy effect feedback matrix, and the policy effect feedback matrix is added to the user feedback data.

[0061] In the embodiments of the present application, after performing the personalized recommendation, data capture is first performed using the buried point technology to record the actual response of users to the recommended content. The buried point technology automatically collects behavior data when users interact with the recommended content by implanting event listening codes in the display interface of the recommended content. When the recommended content is displayed, a click event listener is added to the front-end page (such as a web page, APP interface) to record whether the user clicks on the recommended content; a scroll listener is set in the content area to capture whether the user browses the recommended content and calculate the scroll depth; for the stay duration, a timer is set when the page is loaded to record the time when the user enters the page, and the total stay duration is calculated when the user leaves the page. For the recommended content on the server side, all behavior data of the user after the recommendation interface request is recorded through log analysis methods and stored in the database to form a user response metric data set.

[0062] After obtaining user response metrics such as click-through rate, conversion rate, and dwell time of users, perform strategy execution tracking according to the time series analysis method to evaluate the execution effects of different recommendation strategies. Specifically, first segment the user behavior data according to fixed time windows (such as 1 hour, 1 day, 7 days) to analyze the performance of the recommended content in different time periods. Use the sliding window method to process the click data of users to observe the short-term and long-term effects of the recommendation strategy. For example, when calculating the click-through rate within a 1-hour window, the formula is where N click is the number of clicks within a certain time window, and N impression is the number of displays of the recommended content. Repeat the calculation of this metric on different time windows to identify whether the recommendation strategy maintains a stable effect in the short term or long term. For the conversion rate, through the user path tracking method, monitor whether the user enters the conversion page (such as the purchase page or registration page) after clicking on the recommended content; if the user finally completes the target behavior, record a successful conversion, and the calculation formula is where N conversion is the number of users who complete the target behavior after clicking on the recommended content. By storing the user behavior logs, track the complete path of the user from the recommended content to the conversion page to evaluate the effectiveness of the recommendation strategy.

[0063] After the strategy execution tracking is completed, use the data mapping method to construct a strategy effect feedback matrix to quantify the actual effects of different recommendation strategies and provide a basis for optimization. The strategy effect feedback matrix adopts the vector representation method, with different recommendation strategies as the rows of the matrix and user response metrics (click-through rate, conversion rate, dwell time) as the columns. The values in the matrix represent the performance of different strategies in different user groups. For example, a certain row in the matrix may correspond to the "personalized recommendation strategy based on user interests", and the values in its different columns represent the CTR, CVR, and dwell time of this strategy in different user groups. Through normalization methods (such as max-min normalization), transform each metric to keep the data ranges of different metrics consistent for subsequent calculations.

[0064] Finally, add the strategy effect feedback matrix to the user feedback data and store it in the database for subsequent optimization of the personalized recommendation strategy. Use the matrix factorization method to perform dimensionality reduction processing on the user feedback data to discover the behavior patterns of different user groups. For example, if a certain recommendation strategy has a high CTR in a certain type of user (such as users who prefer digital products), automatically increase the weight of this strategy to give priority to this strategy in future recommendations. At the same time, use stream data processing technology to enable the user feedback data to be collected and calculated in real time, ensuring that the recommendation system can quickly respond to changes in user interests and provide more accurate personalized recommendations.

[0065] Further, in the method provided by the application embodiment, in response to updating the personalized recommendation according to the user feedback data to obtain a behavior analysis result, it further includes:

[0066] Performing reinforcement learning on the user feedback data according to the policy effect feedback matrix to generate a learning result; performing confidence analysis on the personalized recommendation based on the learning result to construct multiple confidence intervals; traversing the multiple confidence intervals according to the personalized recommendation for comparison to generate a data group to be updated; using the data group to be updated as an identifier to perform behavior impact analysis on the personalized recommendation to generate an impact coefficient, and responding to the personalized recommendation according to the impact coefficient to obtain a behavior analysis result.

[0067] In the application embodiment, when performing reinforcement learning on user feedback data according to the policy effect feedback matrix, a Markov decision process (MDP) model is first constructed, where the state space is dynamic customer portrait information, the action space is the policy effect feedback matrix, and the reward function is the weighted score of the user feedback data. The MDP model is used to describe the change of the user state under different recommendation policies, and through iterative optimization, the optimal policy decision path is determined. The value iteration method is adopted to update the state value through the Bellman equation to optimize the recommendation policy. Finally, the learning result is generated by matching the policy decision path with the data learning path.

[0068] After generating the learning result, confidence analysis is performed on the personalized recommendation based on the learning result, and multiple confidence intervals are constructed to measure the reliability of different recommendation policies. The Bootstrap sampling method is adopted, and by randomly sampling the performance of the recommendation policies for different user groups multiple times, the mean click-through rate, mean conversion rate, mean stay duration and the corresponding standard deviation of the policy are calculated, so as to construct the confidence interval. The confidence interval calculation formula is where CI is the confidence interval, is the mean of the click-through rate, conversion rate or stay duration, Z α / 2 is the quantile of the standard normal distribution, σ is the standard deviation of the click-through rate, conversion rate or stay duration, and n is the sample size. By constructing multiple confidence intervals, the reliability of different recommendation policies is quantified to ensure that the recommended content has a high credibility in a statistical sense. For example, if the mean click-through rate of a certain recommendation policy is high but the standard deviation is large, it indicates that the performance of this policy fluctuates greatly, and the recommendation system needs to adjust its stability; if both the mean and standard deviation of the conversion rate are low, it indicates that this policy has poor effect in user conversion and needs to be optimized or replaced.

[0069] After constructing the confidence interval, traverse all personalized recommendation strategies and generate a data group to be updated based on the comparative analysis of the confidence interval. Adopt the confidence interval cross-validation method to compare the confidence intervals of each strategy. If the confidence intervals of the click-through rate, conversion rate, and dwell time of a certain recommendation strategy are significantly higher than those of other strategies, then this strategy is preferentially retained; if the confidence interval of a certain strategy has a large overlap with other strategies and the mean value is low, then this strategy may have no significant advantage and needs to be optimized or replaced. For example, in the A / B test scenario, if the confidence interval of the click-through rate of strategy A is [0.10, 0.15] and the confidence interval of the click-through rate of strategy B is [0.14, 0.18], then it can be determined that strategy B has a higher recommendation effect, and mark strategy A as the data group to be updated.

[0070] After generating the data group to be updated, conduct a behavioral impact analysis on the personalized recommendation based on this data group, calculate the impact coefficient of the recommendation strategy, and finally optimize and adjust the personalized recommendation. Specifically, first perform outlier detection on the data group to be updated to identify abnormal user interaction patterns, and trigger warning signals according to the abnormal interaction patterns. Subsequently, conduct a behavioral impact analysis on the personalized recommendation according to the warning signals, calculate the impact coefficient, and traverse the preset strategy space based on this impact coefficient to determine the optimal alternative strategy. Finally, adjust the personalized recommendation content according to the alternative strategy, optimize the recommendation effect, and generate the final behavioral analysis result.

[0071] Furthermore, in the method provided by the application embodiment, when performing reinforcement learning on the user feedback data according to the strategy effect feedback matrix to generate a learning result, it further includes:

[0072] Define the state space as the dynamic customer portrait information, define the action space as the strategy effect feedback matrix, and the reward function as the weighted score of the user feedback data, and construct a Markov decision process model; perform iterative optimization on the user feedback data through the Markov decision process model to determine the strategy decision path; match according to the strategy decision path and the data learning path to generate the learning result.

[0073] In the embodiment of the present application, first define the state space as the dynamic customer portrait information, which includes features such as the user's browsing history, click behavior, purchase record, interest preference, etc. Adopt the feature engineering method to extract features from the user behavior data and construct a high-dimensional feature vector according to the behavior category. Specifically, perform time series segmentation on the user's historical interaction data. For example, store the browsing, clicking, purchasing, etc. behaviors in the recent 7 days, 30 days, and 90 days as features under different time windows respectively. Use a normalization method (such as Min-Max normalization) to standardize the features to ensure that feature values of different scales can be uniformly used for calculation.

[0074] After the construction of the state space is completed, the action space is defined as the policy effect feedback matrix, which contains the execution methods of different recommendation policies, such as collaborative filtering-based recommendation, content-based recommendation, or rule-based recommendation, etc. The matrix factorization method (such as singular value decomposition SVD) is used to reduce the dimension of the user-recommendation policy interaction data and extract the most important recommendation policy factors. For example, assume that a user has different interaction effects with different types of recommendation policies (based on historical browsing, based on similar users, based on popular products, etc.) over a period of time. The contribution degree of each policy is calculated through matrix factorization and used as the input of the action space to select the optimal policy in the subsequent optimization process.

[0075] To measure the effectiveness of different recommendation policies, a reward function is defined, and the reward value is calculated through linear regression. The goal of the reward function is to quantify the user feedback and guide the reinforcement learning process. First, key indicators such as click-through rate (CTR), conversion rate (CVR), and dwell time (TDR) are extracted from the user feedback data, and the weighted scores of these indicators are calculated using a linear regression model. The formula is R(s, a) = T1·CTR + T2·CVR + T3·TDR, where T1, T2, and T3 are the weights of different feedback indicators, which are obtained by training with historical data to ensure that they can best reflect the user's acceptance of the recommended content. The linear regression model is trained using the least squares method to minimize the error between the predicted reward value and the true user feedback.

[0076] After constructing the Markov decision process model, the Q-learning method is used to learn from the user feedback data to optimize the recommendation policy selection process. Q-learning executes different recommendation policies in different states and adjusts the scores according to the user feedback, enabling the system to find the policy with the highest long-term benefit. Specifically, first, the policy score matrix is initialized, and initial scores are assigned to the recommendation policies in different states. Subsequently, after each round of recommendation execution, the recommendation results are analyzed, and the policy scores are updated according to the feedback data such as the user's clicks, conversions, and dwell time. During the update process, the recommendation policies with better historical performance are preferentially selected, and new recommendation policies are tried with a certain probability to explore potential better solutions. As the learning process progresses, the policy selection is gradually optimized to tend towards the long-term optimal recommendation plan. Through this step, the long-term optimal recommendation policy is determined, and the recommendation decisions in different states become more adaptable.

[0077] During the policy optimization process, a greedy policy is adopted to determine the policy decision path. That is, at each step of the recommendation decision, the recommendation policy with the highest current score is preferentially selected, and the policy execution method is adjusted when necessary. For example, if the conversion rate of a certain recommendation policy is low in multiple past time steps, the score of this policy is reduced, and other policies with higher scores are preferentially selected in the next recommendation. In addition, the state transition process of each decision step is recorded, and the policy selection is continuously adjusted based on the long-term behavior feedback of users, enabling the system to gradually converge to the optimal decision path. Through this step, a policy execution path that can maximize the long-term benefits of users is obtained.

[0078] After completing the policy optimization, the policy decision path is matched with the data learning path to generate the final learning result. The data learning path is the actual interaction trajectory of users, such as clicking on recommended content, browsing details, purchasing, or bouncing out. The dynamic time warping (DTW) method is used to match the policy decision path with the user behavior data to evaluate the consistency between the recommendation policy and the user behavior. The DTW method calculates the minimum cumulative distance between sequences at different time steps, ensuring that even if there is a time deviation between the user behavior and the recommendation policy, the impact of the policy on the user behavior can still be evaluated. For example, if the click-through rate of a certain recommendation policy shows a strong impact after 3 days instead of taking effect immediately, DTW can capture this delay effect and optimize the policy accordingly.

[0079] Finally, the learning result is generated based on the matching result, which includes the optimal recommendation policy, the cumulative reward value of the recommendation policy, and the applicability of different recommendation policies in the user group.

[0080] Furthermore, in the method provided by the application embodiment, using the to-be-updated data group as an identifier, the behavior impact analysis of the personalized recommendation is performed to generate an impact coefficient, and the personalized recommendation is responded according to the impact coefficient to obtain the behavior analysis result, which also includes:

[0081] Outlier detection is performed based on the identifier of the to-be-updated data group to identify abnormal interaction patterns; a warning signal is triggered according to the abnormal interaction patterns, and the behavior impact analysis of the personalized recommendation is performed according to the warning signal to generate an impact coefficient; alternative policies are determined by traversing the preset policy space based on the impact coefficient; and the behavior analysis result is generated by responding to the personalized recommendation according to the alternative policies.

[0082] In the embodiments of the present application, outlier detection is performed based on the identifiers of the data groups to be updated to identify abnormal interaction patterns and optimize personalized recommendation strategies. First, the Isolation Forest algorithm is used for anomaly detection. This method constructs multiple random decision trees, partitions the user behavior data, and calculates the anomaly scores of each user. Behavioral data is extracted from features such as the number of clicks, conversion rate, and dwell time of users and input into the Isolation Forest model for analysis. If the anomaly score of a certain user is higher than the set threshold, for example, a large number of clicks in a short period of time but no conversion, or quickly skipping all recommended content in a short period of time, then this user is marked as a user with abnormal behavior. Among them, the anomaly score is calculated by the Isolation Forest algorithm, Calculated, where E(h(x)) is the average path length of this data point in the forest, and c(n) is a constant used for normalization, depending on the data scale n. If the anomaly score is greater than a certain set threshold (such as 0.8), it is considered that the behavior pattern of this user is significantly different from the overall user group. Through this step, abnormal interaction patterns are identified, providing a basis for subsequent adjustment of the recommendation strategy.

[0083] After the outlier detection is completed, a warning signal is triggered based on the identified abnormal interaction pattern to ensure that the recommendation system can be dynamically adjusted in case of anomalies. The fixed threshold method is set, that is, if the anomaly score of a certain user exceeds the preset risk level (such as 0.8), a warning signal is automatically triggered. For example, if the anomaly scores of a certain recommendation strategy are generally high among multiple user groups, it is marked that there may be problems with this strategy, such as the recommended content not matching or users not being interested in it. The trigger threshold of the warning signal is set to the 95th percentile of the historical user anomaly scores. This can ensure that only the most extreme abnormal behaviors will trigger system alarms. In this step, a warning signal is generated, enabling the system to adjust the recommendation strategy in a timely manner and avoiding unreasonable recommended content from affecting the user experience.

[0084] After the warning signal is triggered, a behavioral impact analysis is performed on the personalized recommendation according to this warning signal to calculate the impact coefficient, which is used to measure the impact of abnormal behavior on the recommendation system. The comparative analysis method is adopted to compare the changes in click-through rate, conversion rate, and dwell time of normal users and abnormal users under the same recommendation strategy. For example, if the proportion of abnormal users of a certain recommendation strategy is higher than that of normal users, and the conversion rate of this strategy is significantly lower, it indicates that this strategy may have a negative impact on the user experience. The impact coefficient is calculated as follows where E[Y|T = 1] represents the mean feedback of abnormal users under a specific recommendation strategy, E[Y|T = 0] represents the mean feedback of normal users, and I represents the degree of influence of this strategy on abnormal behavior. If I is much greater than 0, it indicates that this strategy may trigger abnormal behavior and its application weight will be reduced. Through this step, the impact coefficients of each recommendation strategy are calculated, and it is judged which strategies may lead to higher abnormal interaction risks.

[0085] After obtaining the influence coefficients, traverse the preset policy space based on these influence coefficients to determine suitable alternative policies for optimizing the recommendation effect. Adopt a simple sorting method to sort all the recommendation policies from low to high according to the influence coefficients, and preferentially select the policies with lower influence coefficients (i.e., lower abnormal risks). For example, if the influence coefficient of the current policy A is high among the abnormal user group while the influence coefficient of policy B is low, then policy B will be preferentially adopted in the next round of recommendation. The sorting of the influence coefficients adopts the minimum influence principle, that is, the optimal policy = argminI, where I is the influence coefficient calculated as described above. In the case where the influence coefficients are similar, further consider factors such as the click-through rate and conversion rate of this policy in the historical data to select the optimal alternative policy. Through this step, alternative policies with lower abnormal risks are successfully selected, providing a better solution for optimizing personalized recommendations.

[0086] Finally, respond to the personalized recommendation according to the alternative policy and generate the behavior analysis result. The system adopts the A / B test method to evaluate the optimized recommendation policy, comparing the click-through rate, conversion rate, and dwell time before and after optimization to ensure that the adjusted recommendation policy can effectively reduce abnormal interaction behaviors and improve the user experience. The evaluation criteria for the A / B test are: where, CTR 新 and CTR 旧 are the click-through rates of the new policy and the old policy respectively, and CVR 新 and CVR 旧 are the conversion rates of the new policy and the old policy respectively. If the optimized policy performs better than the original policy in multiple user groups, then further promote this policy and continuously optimize it in future recommendation processes. Through this step, the behavior analysis result is finally generated, enabling the personalized recommendation system to more accurately adapt to the changes in user interests, while reducing the negative impacts brought by abnormal interactions and improving the stability and reliability of the recommendation system.

[0087] In the embodiments of this application, in summary, the embodiments of this application have at least the following technical effects:

[0088] This application collects multi-channel user behavior data in real time to construct dynamic customer portrait information; conducts multi-dimensional predictions based on the dynamic customer portrait information to generate user behavior prediction results; automatically adjusts the promotion strategy according to the AI-driven user behavior prediction results to generate personalized recommendations; executes the personalized recommendations for real-time tracking to generate user feedback data, and updates and responds to the personalized recommendations according to the user feedback data to obtain behavior analysis results. The present invention solves the technical problem in the prior art that the user behavior prediction is inaccurate, resulting in a low matching degree of the promotion strategy. By collecting multi-channel user behavior data in real time, constructing a dynamic customer portrait, conducting multi-dimensional predictions based on AI, and automatically adjusting the promotion strategy, the technical effect of improving the accuracy of the promotion strategy is achieved.

[0089] Embodiment 2, based on the same inventive concept as the AI-driven real-time customer behavior analysis response method in the foregoing embodiment, as Figure 2 shown, this application provides an AI-driven real-time customer behavior analysis response system. The system in the embodiment of this application and the method embodiment are based on the same inventive concept. Among them, the system includes:

[0090] A portrait information construction module 11, which is used to collect multi-channel user behavior data in real time to construct dynamic customer portrait information; a multi-dimensional prediction module 12, which is used to conduct multi-dimensional predictions based on the dynamic customer portrait information to generate user behavior prediction results; a personalized recommendation generation module 13, which is used to automatically adjust the promotion strategy according to the AI-driven user behavior prediction results to generate personalized recommendations; an update response module 14, which is used to execute the personalized recommendations for real-time tracking to generate user feedback data, and update and respond to the personalized recommendations according to the user feedback data to obtain behavior analysis results.

[0091] Furthermore, the system is also used to implement the following functions:

[0092] Real-time access to the user behavior data source through the API interface for data collection to obtain raw data, where the raw data includes customer initial portrait information; extract features from the raw data to generate a structured behavior feature set, conduct time series analysis based on the structured behavior feature set to generate a behavior feature time series; perform dynamic weight allocation on the structured behavior feature set according to the behavior feature time series, and generate user attribute tags according to the weight allocation result; update the customer initial portrait information based on the user attribute tags to construct the dynamic customer portrait information.

[0093] Furthermore, the system is also used to implement the following functions:

[0094] Perform context analysis based on the behavioral feature time series, determine context - associated data, fuse the context - associated data, and construct a pre - trained intent classification model; synchronize the dynamic customer portrait information to the pre - trained intent classification model to obtain intent probability distribution information; set a probability distribution threshold, traverse the intent probability distribution information according to the probability distribution threshold for screening, and generate high - confidence intent labels; perform identification deduction on the dynamic customer portrait information according to the high - confidence intent labels to generate the user behavior prediction result.

[0095] Furthermore, the system is also used to implement the following functions:

[0096] Construct a preset policy space, make a random selection based on the preset policy space to determine a promotion strategy; use a sliding time window mechanism to perform feature engineering processing on the behavioral feature time series to extract a time - series behavior feature vector; perform user interest decay analysis on the dynamic customer portrait according to the time - series behavior feature vector to generate an incremental portrait update package; classify based on the incremental portrait update package to determine a user intent classification result, map the user intent classification result to the preset policy space for matching to generate a policy matching result; automatically optimize the promotion strategy according to the policy matching result to generate the personalized recommendation.

[0097] Furthermore, the system is also used to implement the following functions:

[0098] Establish a mapping based on the time - series behavior feature vector for user activity to obtain a decay factor, and the formula is:

[0099]

[0100] where α is the decay factor, A is the user activity determined according to the time - series behavior feature vector, and the time - series behavior feature vector is in a direct - proportion relationship with the user activity. A avg is the historical average value of the user activity, and σ A is the standard deviation; update the historical feature weight values of the dynamic customer portrait according to the decay factor to generate multiple feature update weight values; perform differential compression according to the multiple feature update weight values to generate a feature change matrix, and align the feature change matrix according to the timestamp to generate the incremental portrait update package.

[0101] Furthermore, the system is also used to implement the following functions:

[0102] Execute the personalized recommendation for capture, obtain user response metrics, track policy execution according to the user response metrics, and generate a policy effect evaluation report; perform data mapping based on the policy effect evaluation report, construct a policy effect feedback matrix, and add the policy effect feedback matrix to the user feedback data.

[0103] Furthermore, the system is also used to implement the following functions:

[0104] Perform reinforcement learning on the user feedback data according to the policy effect feedback matrix to generate a learning result; perform confidence analysis on the personalized recommendation based on the learning result to construct multiple confidence intervals; compare through the personalized recommendation traversing the multiple confidence intervals to generate a data group to be updated; use the data group to be updated as an identifier to perform behavioral impact analysis on the personalized recommendation to generate an impact coefficient, and respond to the personalized recommendation according to the impact coefficient to obtain a behavioral analysis result.

[0105] Furthermore, the system is also used to implement the following functions:

[0106] Define the state space as the dynamic customer portrait information, define the action space as the policy effect feedback matrix, and the reward function as the weighted score of the user feedback data to construct a Markov decision process model; perform iterative optimization on the user feedback data through the Markov decision process model to determine the policy decision path; match according to the policy decision path and the data learning path to generate the learning result.

[0107] Furthermore, the system is also used to implement the following functions:

[0108] Perform outlier detection based on the identifier of the data group to be updated to identify abnormal interaction patterns; trigger a warning signal according to the abnormal interaction pattern, perform behavioral impact analysis on the personalized recommendation according to the warning signal to generate an impact coefficient; determine alternative policies by traversing a preset policy space based on the impact coefficient; respond to the personalized recommendation according to the alternative policies to generate the behavioral analysis result.

[0109] It should be noted that the above order of the embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. And the above describes specific embodiments of this specification. The processes depicted in the drawings do not necessarily require the specific order and continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0110] The foregoing are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

[0111] This specification and the drawings are merely illustrative of the present application and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application is intended to include these changes and modifications.

Claims

1. AI-driven real-time customer behavior analysis response method, characterized in that, The method includes: Collecting multi-channel user behavior data in real time to construct dynamic customer portrait information; Performing multi-dimensional prediction based on the dynamic customer portrait information to generate user behavior prediction results; Automatically adjusting the promotion strategy according to the AI-driven user behavior prediction results to generate personalized recommendations; Executing the personalized recommendations for real-time tracking to generate user feedback data, and updating and responding to the personalized recommendations according to the user feedback data to obtain behavior analysis results.

2. The AI-driven real-time customer behavior analysis response method according to claim 1, characterized in that The method of collecting multi-channel user behavior data in real time to construct dynamic customer portrait information includes: Real-time accessing user behavior data sources through API interfaces for data collection to obtain raw data, where the raw data contains initial customer portrait information; Performing feature extraction on the raw data to generate a structured behavior feature set, and performing time series analysis based on the structured behavior feature set to generate a behavior feature time series; Performing dynamic weight assignment on the structured behavior feature set according to the behavior feature time series, and generating user attribute tags according to the weight assignment results; Updating the initial customer portrait information based on the user attribute tags to construct the dynamic customer portrait information.

3. The AI-driven real-time customer behavior analysis response method according to claim 2, wherein, The method of performing multi-dimensional prediction based on the dynamic customer portrait information to generate user behavior prediction results includes: Performing context analysis based on the behavior feature time series to determine context-related data, and fusing the context-related data to construct a pre-trained intent classification model; Synchronizing the dynamic customer portrait information to the pre-trained intent classification model to obtain intent probability distribution information; Setting a probability distribution threshold, and traversing and screening the intent probability distribution information according to the probability distribution threshold to generate high-confidence intent tags; Performing identification deduction on the dynamic customer portrait information according to the high-confidence intent tags to generate the user behavior prediction results.

4. The AI-driven real-time customer behavior analysis response method according to claim 2, wherein, The method of automatically adjusting the promotion strategy according to the AI-driven user behavior prediction results to generate personalized recommendations includes: Constructing a preset policy space, randomly selecting based on the preset policy space to determine the promotion strategy; Performing feature engineering processing on the behavior feature time series by using a sliding time window mechanism to extract time series behavior feature vectors; Performing user interest decay analysis on the dynamic customer portrait according to the time series behavior feature vectors to generate an incremental portrait update package; Classifying based on the incremental portrait update package to determine the user intent classification result, mapping the user intent classification result to the preset policy space for matching to generate a policy matching result; Automatically optimizing the promotion strategy according to the policy matching result to generate the personalized recommendation.

5. The AI-driven real-time customer behavior analysis response method according to claim 4, characterized in that, The method of performing user interest decay analysis on the dynamic customer portrait according to the time series behavior feature vectors to generate an incremental portrait update package includes: Establishing a mapping of user activity based on the time series behavior feature vectors to obtain a decay factor, and the formula is: where α is the attenuation factor, A is the user activity determined according to the time series behavior feature vector, the time series behavior feature vector has a positive proportional relationship with the user activity, and A avg is the historical average value of the user activity, and σ A is the standard deviation; Updating the historical feature weight values of the dynamic customer portrait according to the decay factor to generate multiple feature update weight values; Update the weight values according to the multiple features for differential compression, generate a feature change matrix, and align the feature change matrix according to the timestamp to generate the incremental portrait update package.

6. The AI-driven real-time customer behavior analysis response method according to claim 4, wherein Execute the personalized recommendation for real-time tracking to generate user feedback data. The method includes: Execute the personalized recommendation for capture to obtain user response metrics, and perform policy execution tracking according to the user response metrics to generate a policy effect evaluation report; Perform data mapping based on the policy effect evaluation report, construct a policy effect feedback matrix, and add the policy effect feedback matrix to the user feedback data.

7. The AI-driven real-time customer behavior analysis response method according to claim 6, wherein Update the response of the personalized recommendation according to the user feedback data to obtain a behavior analysis result. The method includes: Perform reinforcement learning on the user feedback data according to the policy effect feedback matrix to generate a learning result; Perform confidence analysis on the personalized recommendation based on the learning result to construct multiple confidence intervals; Traverse the multiple confidence intervals according to the personalized recommendation for comparison to generate a data group to be updated; Use the data group to be updated as an identifier to perform behavior impact analysis on the personalized recommendation to generate an impact coefficient, and respond to the personalized recommendation according to the impact coefficient to obtain a behavior analysis result.

8. The AI-driven real-time customer behavior analysis response method according to claim 7, wherein Perform reinforcement learning on the user feedback data according to the policy effect feedback matrix to generate a learning result. The method includes: Define the state space as the dynamic customer portrait information, define the action space as the policy effect feedback matrix, and the reward function as the weighted score of the user feedback data, and construct a Markov decision process model; Iteratively optimize the user feedback data through the Markov decision process model to determine the policy decision path; Match the policy decision path with the data learning path to generate the learning result.

9. The AI-driven real-time customer behavior analysis response method according to claim 7, wherein Use the data group to be updated as an identifier to perform behavior impact analysis on the personalized recommendation to generate an impact coefficient, and respond to the personalized recommendation according to the impact coefficient to obtain a behavior analysis result. The method includes: Perform outlier detection based on the identifier of the data group to be updated to identify abnormal interaction patterns; Trigger a warning signal according to the abnormal interaction pattern, and perform behavior impact analysis on the personalized recommendation according to the warning signal to generate an impact coefficient; Traverse the preset policy space based on the impact coefficient to determine alternative policies; Respond to the personalized recommendation according to the alternative policy to generate the behavior analysis result.

10. An AI-driven real-time customer behavior analysis response system, characterized in that, The system includes: A portrait information construction module for real-time collecting multi-channel user behavior data to construct dynamic customer portrait information; A multi-dimensional prediction module for performing multi-dimensional prediction based on the dynamic customer portrait information to generate a user behavior prediction result; A personalized recommendation generation module for automatically adjusting the promotion strategy according to the AI-driven user behavior prediction result to generate a personalized recommendation; An update response module for executing the personalized recommendation for real-time tracking to generate user feedback data, and updating the response of the personalized recommendation according to the user feedback data to obtain a behavior analysis result.

Images