Recommendation system based on user behavior analysis

Abstract

The invention discloses a recommendation system based on user behavior analysis, and relates to the technical field of artificial intelligence and information recommendation, and the system comprises a front-end signal acquisition module, a layered pulse processing module, a differential signal amplification module and a recommendation strategy generation module. User behaviors are divided into three levels of millisecond-level instantaneous pulse, minute-level behavior wave and hour-level trend flow, and the relevance of behavior mutation of each level is extracted through a differential signal amplifier, so that the system can sensitively capture behavior changes of a user under different time scales, accurately insight into instant interests and long-term demands of the user, and improve the user experience. The system can quickly find the attention of a user to specific contents through millisecond-level instantaneous pulse analysis, and further understand the behavior preference and potential requirements of the user by combining the analysis of minute-level behavior waves and hour-level trend flow.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and information recommendation, and specifically to a recommendation system based on user behavior analysis. Background Art

[0002] In the digital age, the Internet has generated a vast amount of information. Various websites and applications provide users with rich and diverse content and services. However, in the face of such a huge information resource, it is often difficult for users to quickly find the information that meets their own needs. The user behavior analysis recommendation system came into being. It collects and analyzes various behavior data of users during use, such as browsing records, click behaviors, search keywords, etc., to understand the interests and needs of users, so as to provide personalized recommendation services for users. Such a recommendation system can not only improve the user experience, help users save time and energy, but also bring higher conversion rates and revenues to enterprises.

[0003] Currently, existing recommendation systems have significant defects when facing complex and changeable user behaviors. In the behavior analysis in the time dimension, most systems lack refined processing of behaviors at different time scales. They are difficult to distinguish behavior signals with different time characteristics such as millisecond-level instantaneous pulses, minute-level behavior waves, and hour-level trend flows, resulting in inaccurate grasping of users' immediate interests and long-term behavior trends. For example, it is impossible to timely capture the interest signals contained in the instantaneous behavior of a user's short pause when scrolling the page, missing the opportunity for accurate recommendation; at the same time, when analyzing the long-term behavior trends of users, it is also impossible to effectively integrate behavior data at multiple time scales, resulting in the recommendation strategy lagging behind the changes in users' actual needs, affecting the overall performance and recommendation effect of the recommendation system.

[0004] In summary, the existing user behavior analysis recommendation systems have obvious deficiencies in the refinement degree of behavior analysis in the time dimension and cannot meet the needs of users for personalized and accurate recommendation services. Therefore, it is of great practical significance to develop a recommendation system that can perform refined analysis of user behaviors at multiple time scales and achieve efficient coordination of each module. Summary of the Invention

[0005] The purpose of the present invention is to make up for the deficiencies of the existing technology and provide a recommendation system based on user behavior analysis. It can divide user behaviors into three levels: millisecond-level instantaneous pulses, minute-level behavior waves, and hour-level trend flows through the constructed phased behavior pulse real-time capture engine, and extract the correlation of behavior mutations at each layer through the differential signal amplification module. This enables the system to keenly capture the behavior changes of users at different time scales, accurately understand the immediate interests and long-term needs of users, so as to push the products or services that best meet their needs to users in the first time, realizing real-time and accurate recommendation, and greatly improving the user experience.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a recommendation system based on user behavior analysis, which consists of: a front-end signal acquisition module, a hierarchical pulse processing module, a differential signal amplification module, and a recommendation strategy generation module;

[0007] The front-end signal acquisition module is used to capture the millisecond-level operation event stream of user interaction in real time, and generate a pulse density matrix through a signal dimensionality reduction unit;

[0008] The hierarchical pulse processing module consists of an instantaneous pulse layer, a behavior wave layer, and a trend flow layer. Among them, the instantaneous pulse layer uses sliding window difference processing for signals in the range of 0.1 - 1000 ms, the behavior wave layer analyzes the behavior entropy change at the 1 - 10 minute level through a dynamic time window, and the trend flow layer uses a time-gated graph neural network to predict the behavior migration path at the 1 - 24 hour level;

[0009] The differential signal amplification module is connected to each signal processing layer and calculates the cross-layer mutual information gain, and generates a recommendation trigger instruction when the gain value exceeds a preset threshold;

[0010] After receiving the trigger instruction, the recommendation strategy generation module combines the user's real-time behavior context to generate a multi-modal recommendation scheme including main promotion content, auxiliary information, and interaction guidance.

[0011] Further, the front-end signal acquisition module includes: a multi-source sensor array, a signal dimensionality reduction unit, and an event trigger logic control unit, where:

[0012] The multi-source sensor array consists of a touch pressure sensor, a cursor trajectory tracker, and a page scrolling monitor, and is used to collect the original sensor data stream of the page scrolling pause duration, touch click interval, and cursor movement acceleration;

[0013] The signal dimensionality reduction unit is used to divide the original sensor data stream into multiple data blocks according to a time window. Each data block has multiple sampling points. Principal component analysis (PCA) is used to perform dimensionality reduction feature extraction on the data block, and the principal components with a variance contribution rate greater than 85% are retained. The dimensionality-reduced feature vectors are encoded into a pulse density matrix, and the matrix dimension is compressed to 20% - 30% of the original data stream;

[0014] The event trigger logic control unit activates data upload when detecting any of the following conditions:

[0015] The page scrolling pause duration > 300 ms and the acceleration change rate < 0.05 m / s 2 ;

[0016] The continuous touch click interval of 3 times < 150 ms;

[0017] The curvature radius mutation of the cursor movement trajectory > 15° / ms;

[0018] The data upload outputs the original sensor data stream after dimensionality reduction to the hierarchical pulse processing module. After three-layer collaborative analysis, structured features are output, and finally, a recommendation strategy is generated.

[0019] Furthermore, the execution steps of the signal dimensionality reduction unit are as follows:

[0020] The original sensor data stream is segmented into data blocks according to the time window t, and each data block contains 500 - 1000 sampling points;

[0021] Feature extraction is performed on the data block through principal component analysis (PCA), and the data block is set as matrix Q n×p , where n represents the number of samples in the data block, that is, the number of sampling points, and p represents the number of features in the data block. For matrix Q n×p Standardization processing is performed to obtain That is where Q ij refers to the j-th eigenvalue of the i-th sample, μ j is the mean of the j-th feature, and σ j is the standard deviation of the j-th feature;

[0022] Calculate the covariance matrix of the standardized matrix where is the transpose matrix of the standardized matrix , and the covariance matrix S is used to measure the correlation between each feature and the variance of each feature;

[0023] Eigenvalues λ1 ≥ λ2 ≥ … ≥ λ p and the corresponding eigenvectors e1, e2, …, e p are obtained through the covariance matrix S. The eigenvalue λ k reflects the amount of information contained in the k-th principal component, and the eigenvector e k determines the direction of the k-th principal component. The variance contribution rate of the k-th principal component where represents the sum of all eigenvalues, and η k represents the proportion of the k-th principal component in the total amount of information. Retain the first m principal components whose variance contribution rate satisfies That is, select these principal components that can cumulatively explain at least 85% of the data variance to achieve data dimensionality reduction and retain the key information of the data to the greatest extent.

[0024] Furthermore, the hierarchical pulse processing module includes: an instantaneous pulse layer, a behavior wave layer, a trend flow layer, and a cross-layer collaborative control unit, where:​

[0025] The instantaneous pulse layer sets the window width to w where 10 ≤ w ≤ 100 ms and the sliding step to s where 1 ≤ s ≤ 10 ms. For the behavior signal x(t) in the pulse density matrix obtained after processing by the signal dimensionality reduction unit, calculate the first-order difference value Δx of the behavior signal within the window k = x k - x k-1 , when the difference values of three consecutive windows exceed twice the standard deviation of the baseline noise level, i.e., Δx k > 2σ noise , it is marked as an invalid instantaneous pulse event, otherwise it is a valid instantaneous pulse event, and σ noise is the standard deviation of the baseline noise level;

[0026] The behavior wave layer divides the behavior sequence with a 1-minute time unit, and uses a sliding window mechanism to dynamically adjust the window range to T where 1 ≤ T ≤ 10 minutes, and calculates the entropy value within each window where p i represents the occurrence probability of behavior sample i within the window, and n is the number of samples in the data block. When the entropy change difference ΔH = |H2 - H1| > 0.5 between two adjacent window entropies H1 and H2, it triggers a behavior pattern migration and is confirmed as an abnormal behavior operation event, that is, the user's behavior pattern changes from one behavior pattern to another in adjacent time windows, otherwise it is a normal behavior operation event;

[0027] The trend flow layer constructs a time-gated graph neural network, retains the key behavior paths through a gating mechanism, and predicts the behavior migration path for the next 2 hours;

[0028] The cross-layer collaborative control unit, when both the instantaneous pulse layer and the behavior wave layer detect invalid and abnormal events, forces the trend flow layer to start real-time graph structure update and allocates an additional 30% of computing resources.

[0029] Furthermore, the trend flow layer constructs a time-gated graph neural network, where the nodes represent user behavior categories, and the edge weights are jointly determined by the behavior transfer frequency and temporal correlation. That is, for the behavior transfer frequency of f ij , and the temporal correlation of r ij , then the weight w of edge (i,j) ij = αf ij + (1 - α)r ij , α is the weight coefficient and 0 < α < 1. Within the range of 1 - 24 hours, the graph structure is dynamically updated every 30 minutes. Through the gating mechanism, the key behavior paths with weight w ij > 0.7 are retained, and when predicting the behavior migration path for the next 2 hours, the historical trend weight w history60% of w and 40% of the real-time behavior entropy change value ΔH generate a mixed attention coefficient β = 0.6w history + 0.4ΔH, and the attention coefficient is used to adjust the attention degree to historical trends and real-time behavior changes when predicting the behavior migration path.

[0030] Furthermore, the differential signal amplification module includes: a cross-layer information collection unit, a mutual information gain calculation unit, and a trigger instruction generation unit, where:

[0031] The cross-layer information collection unit collects the processed signal feature data of each layer from the instantaneous pulse layer, the behavior wave layer, and the trend flow layer. For the instantaneous pulse layer, it collects the signal parameters marked as valid instantaneous pulse events. For the behavior wave layer, it collects the entropy change difference and the behavior pattern migration information. For the trend flow layer, it collects the node information, edge weights, and mixed attention coefficients of the time-gated graph neural network;

[0032] The mutual information gain calculation unit calculates the mutual information between the signal features of each layer, which is used to analyze the mutual information gain between user behavior features and recommended content, and judge the feedback degree of different user behaviors to different recommended content;

[0033] The trigger instruction generation unit generates a recommendation trigger instruction when the mutual information gain values of at least two of the three-layer behavior signals exceed a preset threshold T g At this time, a recommendation trigger instruction is generated, and this instruction contains the association information of the signal features of each layer.

[0034] Furthermore, the mutual information gain calculation unit sets the signal feature set of the instantaneous pulse layer as X, the signal feature set of the behavior wave layer as Y, and the signal feature set of the trend flow layer as Z. For X and Y, the mutual information where p(x,y) is the joint occurrence probability of X and Y within the window, and p(x) and p(y) are the marginal probability distributions of X and Y respectively. Similarly, I(X;Z) and I(Y;Z) are obtained. At time t and t + 1 of the time window T, by quantifying the change amount of the mutual information between different layers, the cross-layer mutual information gain G is obtained, that is, for the mutual information gain G of X and Y layers XY (t,t + 1) = I(X;Y) t+1 -I(X;Y) t , and similarly G XZ and G YZ .

[0035] Furthermore, the recommendation strategy generator module includes: a context analysis unit, a recommended content screening unit, an interaction guidance design unit, and a multi-modal recommendation scheme generation unit, where:

[0036] The context analysis unit receives the recommended trigger instruction sent by the differential signal amplifier module and analyzes it in combination with the user's real-time behavior context information, where the real-time behavior context information includes the page the user is currently on, browsing history, search records, and the device used;

[0037] The recommended content screening unit filters out the main recommended content that meets the user's needs from the content library according to the context analysis result. At the same time, relevant auxiliary information is screened out for each piece of main recommended content;

[0038] The interaction guidance design unit designs the interaction guidance method according to the main recommended content and the auxiliary information, and the interaction guidance method includes button prompts, pop-up reminders, and recommended list displays;

[0039] The multi-modal recommendation scheme generation unit integrates the main recommended content, the auxiliary information, and the interaction guidance method into a multi-modal recommendation scheme, and the multi-modal recommendation scheme is presented to the user in various forms such as text, pictures, videos, and audio.

[0040] Compared with the prior art, the recommendation system based on user behavior analysis has the following beneficial effects:

[0041] First, the phased behavior pulse real-time capture engine constructed in the present invention divides user behavior into three levels: millisecond-level instantaneous pulses, minute-level behavior waves, and hour-level trend flows, and extracts the correlation of behavior mutations at each level through a differential signal amplifier. This enables the system to keenly capture the behavior changes of users at different time scales, accurately understand the user's immediate interests and long-term needs, so that the system can quickly discover the user's attention to specific content through millisecond-level instantaneous pulse analysis, and further understand the user's behavior preferences and potential needs by combining the analysis of minute-level behavior waves and hour-level trend flows, thereby pushing the products and services that best meet the user's needs to the user in the first time, realizing real-time and accurate recommendations, and greatly improving the user experience.

[0042] Second, the front-end signal acquisition module, the hierarchical pulse processing module, and the differential signal amplification module in the present invention can cooperate closely to achieve rapid information transmission and sharing, dynamically allocate hardware resources according to the real-time pulse density, and ensure the stable operation of the system under high load. This module cooperation and resource optimization mechanism not only improves the operation efficiency of the recommendation system, but also enhances the scalability and adaptability of the system, enabling it to better handle the processing requirements of user behavior data with different scales and complexities, and providing strong support for the application of the recommendation system in more fields.

[0043] Other advantages, objects, and features of the present invention will be set forth in part in the following description, and in part will be obvious to those skilled in the art from a review of the following, or may be learned from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0045] Figure 1 is an operation flowchart of a recommendation system based on user behavior analysis;

[0046] Figure 2 is a schematic diagram of the module composition of a recommendation system based on user behavior analysis. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will, in conjunction with the drawings and preferred embodiments, describe in detail the specific embodiments, structures, features, and effects of the present invention as follows.

[0048] Embodiment 1

[0049] This embodiment elaborates in detail the recommendation process of a recommendation system based on user behavior analysis when a user starts using a shopping platform to browse products. As Figure 2 shown, the system collects user behavior data through a front-end signal acquisition module, generates a pulse density matrix through processing such as signal dimensionality reduction, the hierarchical pulse processing module analyzes the behavior signals from different time scales, the differential signal amplification module calculates the mutual information gain to trigger recommendations, and the recommendation strategy generation module generates a multi-modal recommendation plan based on the trigger instruction. Each module works together to accurately understand the user's needs and achieve personalized recommendations.

[0050] In a specific implementation, the front-end signal acquisition module includes: a multi-source sensor array, a signal dimension reduction unit, and an event triggering logic control unit, which are used to capture the millisecond-level operation event stream of user interaction in real time and generate a pulse density matrix through the signal dimension reduction unit. The multi-source sensor array is responsible for collecting different types of user behavior data through a touch pressure sensor, a cursor tracker, and a page scrolling monitor. The touch pressure sensor collects pressure change data when the user touches the screen in real time, and its collection frequency can ensure accurate recording of each touch operation. The cursor tracker records the movement trajectory of the cursor during the user operation, including information such as movement acceleration. The page scrolling monitor monitors the page scrolling situation at all times and records the page scrolling pause duration. These sensors collect raw sensor data streams, which contain a large amount of user behavior information, but the data is relatively messy and needs to be further processed by the signal dimension reduction unit. The signal dimension reduction unit first divides the raw sensor data stream into time windows, and each time window forms a data block. Each data block contains a certain number of sampling points. Let the data block be the matrix Q n×p , where n represents the number of samples in the data block (i.e., the number of sampling points), and p represents the number of features in the data block. In this embodiment, since the collected behavioral data includes multiple types such as touch pressure, cursor movement, and page scrolling, p is the sum of the number of features corresponding to these different types of data. n×p Standardization is performed to eliminate the impact of dimension and data distribution differences between different features, making subsequent analysis more accurate. Where Q ij Refers to the jth eigenvalue of the i-th sample, μ j is the mean of the jth feature, σ j is the standard deviation of the jth feature, and each eigenvalue is standardized by standardization to obtain a standardized matrix Then calculate the normalized matrix The covariance matrix of in is the standardized matrix The covariance matrix S is used to measure the correlation between each feature and the variance of each feature. It can reflect the relationship between different behavioral features. The covariance matrix S can be used to obtain the eigenvalues λ1≥λ2≥…≥λ p and the corresponding eigenvectors e1,e2,…,e p , eigenvalue λ k Reflects the amount of information contained in the kth principal component, the eigenvector e k The direction of the kth principal component is determined, and the variance contribution rate of the kth principal component is here represents the sum of all eigenvalues, ηk which represents the proportion of the k-th principal component in the total information volume. To achieve data dimensionality reduction and retain the key information of the data to the greatest extent, the signal dimensionality reduction unit retains the first m principal components whose variance contribution rate satisfies , encodes the dimensionality-reduced feature vector into a pulse density matrix, and compresses the matrix dimension to 20%-30% of the original data stream, which not only reduces the amount of data but also retains the information valuable for subsequent analysis; the event-triggered logic control unit constantly monitors the collected user behavior data. When it detects that the page scrolling pause duration exceeds a certain time length (set as t1, corresponding to the page scrolling pause duration > 300 ms in the document in this embodiment) and the acceleration change rate is less than a certain value (set as a1, corresponding to the acceleration change rate < 0.05 m / s 2 ), or the interval between three consecutive touch clicks is less than a certain time (set as t2, corresponding to the interval between three consecutive touch clicks < 150 ms), or the curvature radius mutation of the cursor movement trajectory exceeds a certain angle (set as θ, corresponding to the curvature radius mutation of the cursor movement trajectory > 15° / ms), the data upload operation is activated, and the data upload outputs the original sensor data stream to the hierarchical pulse processing module.

[0051] The hierarchical pulse processing module includes: an instantaneous pulse layer, a behavior wave layer, a trend flow layer, and a cross-layer collaborative control unit. Among them, the instantaneous pulse layer receives the behavior signal x(t) in the pulse density matrix processed by the front-end signal acquisition module. To capture the rapid changes of the behavior signal, the instantaneous pulse layer sets the window width as w (10 ≤ w ≤ 100 ms) and the sliding step as s (1 ≤ s ≤ 10 ms), and uses the sliding window difference algorithm to calculate the first-order difference value of the behavior signal within the window. For the behavior signal within the window, assume the signal value at the k-th time point is x k , then the first-order difference value Δx k = x k - x k-1 . By calculating the difference value, the change situation of the behavior signal can be highlighted. For example, when there is a sudden change in the instantaneous behavior during the user's operation, the instantaneous pulse layer will compare the difference values of three consecutive windows with twice the standard deviation of the baseline noise level (assume the standard deviation of the baseline noise level is σ noise ). When the difference values of three consecutive windows exceed 2σ noiseWhen it is, it is marked as an invalid instantaneous pulse event, which means that a significant change in the user's behavior has been detected, which may represent that the user has paid special attention to a certain content or performed an abnormal operation. Otherwise, it is a valid instantaneous pulse event. The marking of the valid instantaneous pulse event will be used as an important basis for subsequent analysis to judge the user's immediate interest points. The behavior wave layer divides the behavior sequence with 1 minute as the time unit and uses a sliding window mechanism to dynamically adjust the window range to T (1 ≤ T ≤ 10 minutes). Within each window, the behavior wave layer calculates the entropy value H. The entropy value is used to measure the uncertainty or chaos degree of the behavior. The calculation formula is where p i represents the occurrence probability of behavior sample i within the window, and n is the number of samples in the data block. By calculating the entropy value, we can understand the stability and changes of the user's behavior over a period of time. When the entropy change difference ΔH = |H2 - H1| between the entropies H1 and H2 of two adjacent windows is > 0.5, a behavior pattern migration is triggered, which indicates that the user's behavior pattern has changed within adjacent time windows, changing from one behavior pattern to another, and it is confirmed as an abnormal behavior operation event. Otherwise, it is an abnormal behavior operation event. For example, if the user mainly browses electronic products within one window and suddenly browses a large number of clothing products in the next window, the entropy change difference will exceed the threshold and trigger a behavior pattern migration. The trend flow layer constructs a time-gated graph neural network. In this network, the nodes represent the user behavior categories, and the edge weights are jointly determined by the behavior transfer frequency and temporal correlation. For the behavior transfer frequency of f ij , and the temporal correlation of r ij , the weight w ij of the edge (i, j) = αf ij + (1 - α)r ij , where α is the weight coefficient and 0 < α < 1, comprehensively considering the frequency of behavior transfer and the temporal precedence relationship, enabling the edge weights to more accurately reflect the relationship between user behaviors. Within the range of 1 - 24 hours, the trend flow layer performs a graph structure dynamic update every 30 minutes. Through the gating mechanism, the key behavior paths with weights w ij > 0.7 are retained. This can timely reflect the changes in the user's behavior pattern, remove unimportant behavior associations, and focus more on the key behavior trends. When predicting the behavior migration path in the next 2 hours, the trend flow layer fuses 60% of the historical trend weight w history and 40% of the real-time behavior entropy change value ΔH to generate a mixed attention coefficient β = 0.6w history+0.4ΔH, which adjusts the attention degree to historical trends and real-time behavior changes when predicting the behavior migration path through this attention coefficient, so as to more accurately predict the future behavior trend of users; the cross-layer collaborative control unit is responsible for monitoring the detection results of the instantaneous pulse layer and the behavior wave layer. When the instantaneous pulse layer and the behavior wave layer simultaneously detect invalid and abnormal events, the cross-layer collaborative control unit forces the trend flow layer to start real-time graph structure update and allocates an additional 30% of computing resources. This is because when this situation occurs, it means that the user's behavior may have changed greatly, and the trend flow layer needs to adjust the analysis and prediction of the user's behavior in a timely manner. Allocating additional computing resources can ensure that the trend flow layer can complete the update operation quickly and accurately, ensuring the system's timely response to changes in the user's behavior.

[0052] The differential signal amplification module includes: a cross-layer information collection unit, a mutual information gain calculation unit, and a trigger instruction generation unit, which are used to connect each signal processing layer and calculate the cross-layer mutual information gain, and generate a recommended trigger instruction when the gain value exceeds a preset threshold. Among them, the cross-layer information collection unit collects the signal feature data processed by each layer from the instantaneous pulse layer, the behavior wave layer, and the trend flow layer, that is, collects the signal parameters marked as valid instantaneous pulse events from the instantaneous pulse layer, such as difference values, occurrence times, etc., collects the entropy change difference and behavior pattern migration information from the behavior wave layer, and collects the node information, edge weights, and mixed attention coefficients of the time-gated graph neural network from the trend flow layer. These data contain the key features of the user's behavior at different time scales and provide a data basis for subsequent calculation of the mutual information gain; the mutual information gain calculation unit sets the signal feature set of the instantaneous pulse layer as X, the signal feature set of the behavior wave layer as Y, and the signal feature set of the trend flow layer as Z. For X and Y, the mutual information where p(x,y) is the joint occurrence probability of X and Y within the window, and p(x) and p(y) are the marginal probability distributions of X and Y respectively. Similarly, I(X;Z) and I(Y;Z) can be obtained. Mutual information is used to measure the degree of dependence between two random variables, which represents the degree of association between behavior characteristics of different layers here. At time t and t + 1 in the time window T, by quantifying the change amount of the mutual information between different layers, the cross-layer mutual information gain G is obtained. For example, for the mutual information gain G of layers X and Y XY (t,t + 1)=I(X;Y) t+1 -I(X;Y) t , and similarly, G X Z and G Y Z. By calculating the cross-layer mutual information gain, it is possible to judge the feedback degree of different user behaviors to different recommended contents, as well as the association change situation between user behavior characteristics at different time scales; the trigger instruction generation unit when the mutual information gain values of at least two of the three-layer behavior signals exceed the preset threshold T gWhen a recommendation trigger instruction is generated, this instruction contains the correlation information of signal features at each layer. For example, which layer's signal changes led to the gain exceeding the threshold, and the specific characteristic parameters of these changes. Integrate the user behavior information analyzed at each layer to provide clear guidance for the recommendation strategy generation module.

[0053] The recommendation strategy generation module includes: a context analysis unit, a recommended content screening unit, an interaction guidance design unit, and a multi-modal recommendation scheme generation unit. After receiving the trigger instruction, it combines the user's real-time behavior context to generate a multi-modal recommendation scheme including the main recommended content, auxiliary information, and interaction guidance. Among them, the context analysis unit receives the recommendation trigger instruction sent by the differential signal amplifier module and analyzes it in combination with the user's real-time behavior context information. The real-time behavior context information includes the page where the user is currently located, browsing history, search records, used devices, etc. For example, if the user is currently browsing the electronic product page, has had multiple records of searching for mobile phone-related keywords before, and is using a mobile device, then all these information will be collected and analyzed by the context analysis unit to more comprehensively understand the user's needs and preferences. The recommended content screening unit screens out the main recommended content that meets the user's needs from the content library according to the context analysis results. The content library stores various product information, including electronic products of different brands, models, and prices. The recommended content screening unit will select the product that best meets the user's interest from the content library as the main recommended content based on the user's browsing history, search records, and current page, etc. At the same time, screen out relevant auxiliary information for each main recommended content, such as detailed product parameters, user reviews, usage tutorials, etc. These auxiliary information can help users better understand the main recommended content, thereby improving the user's acceptance of the recommended content; the interaction guidance design unit designs the interaction guidance method according to the main recommended content and auxiliary information. The interaction guidance method includes button prompts, pop-up reminders, recommended list displays, etc., and displays relevant products in the form of a recommended list to provide more options; the multi-modal recommendation scheme generation unit integrates the main recommended content, auxiliary information, and interaction guidance method into a multi-modal recommendation scheme. The multi-modal recommendation scheme is presented to the user in various forms such as text, image, video, and audio to increase the user's attention and understanding of the recommended content, thereby improving the recommendation effect.

[0054] In summary, in this embodiment, the recommendation system based on user behavior analysis realizes the refined analysis of user behavior at multiple time scales and personalized recommendation through the collaborative work of the front-end signal acquisition module, the hierarchical pulse processing module, the differential signal amplification module, and the recommendation strategy generation module. The front-end signal acquisition module collects and processes user behavior data, providing a basis for subsequent analysis; the hierarchical pulse processing module analyzes behavior signals at different time scales to capture changes in user behavior; the differential signal amplification module calculates the mutual information gain to judge the correlation between user behavior and recommended content and triggers recommendations; the recommendation strategy generation module generates multi-modal recommendation schemes according to the trigger instructions to meet user needs, improving the accuracy of recommendations and system performance, and providing a better user experience for users.

[0055] Embodiment 2

[0056] As Figure 1 shown, the present invention realizes accurate personalized recommendation by a recommendation system based on user behavior analysis that captures users' instantaneous operations in real time, analyzes multi-level behavior patterns, and dynamically adjusts recommendation strategies. The specific steps are as follows:

[0057] (1) Real-time collection of user operation signals

[0058] The front-end signal acquisition module continuously monitors user interaction behaviors, including: page scrolling pauses, touch click frequencies, and cursor movement trajectories, ensuring the capture of subtle instantaneous behaviors;

[0059] The signal dimensionality reduction unit compresses a large amount of raw data into key features, divides data blocks according to time windows, and extracts features;

[0060] Generate a lightweight pulse density matrix to activate data upload;

[0061] (2) Hierarchical analysis of user behavior

[0062] The instantaneous pulse layer identifies immediate interests and filters out invalid noises;

[0063] The behavior wave layer; detects changes in behavior patterns, calculates the behavior entropy value, and judges whether the user enters the exploration period;

[0064] The trend flow layer predicts future behavior paths and integrates long-term preferences with real-time scenarios;

[0065] (3) Cross-layer correlation analysis and trigger recommendations

[0066] The differential signal amplification module analyzes the correlation of data in each layer and calculates the cross-layer mutual information gain;

[0067] The trigger instruction generation unit triggers a recommendation instruction when the mutual information gain values of at least two layers exceed a preset threshold;

[0068] If the behavior pattern mutates, trigger cross-layer collaborative adjustment of resources;

[0069] Send a recommendation instruction to the policy generation module;

[0070] (4) Dynamically generate multi-modal recommendations

[0071] The context analysis unit combines the trigger instruction with the user's real-time behavior to screen the recommended content;

[0072] The recommendation policy generation module integrates the main promotion content, auxiliary information, and interactive guidance, and outputs a multi-modal recommendation plan in the form of pictures, videos, or audios.

[0073] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications within the scope of the technical solution of the present invention to obtain equivalent embodiments with equivalent changes. However, as long as it does not depart from the content of the technical solution of the present invention, any brief modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A recommendation system based on user behavior analysis, characterized in that, The system consists of: a front-end signal acquisition module, a hierarchical pulse processing module, a differential signal amplification module, and a recommendation strategy generation module; The front-end signal acquisition module is used to capture the millisecond-level operation event stream of user interaction in real time and generate a pulse density matrix through a signal dimensionality reduction unit; The hierarchical pulse processing module consists of an instantaneous pulse layer, a behavior wave layer, and a trend flow layer. Among them, the instantaneous pulse layer uses sliding window difference processing for signals at the 0.1 - 1000 ms level, the behavior wave layer analyzes the behavior entropy change at the 1 - 10 minute level through a dynamic time window, and the trend flow layer uses a time-gated graph neural network to predict the behavior migration path at the 1 - 24 hour level; The differential signal amplification module is connected to each signal processing layer and calculates the cross-layer mutual information gain. When the gain value exceeds the preset threshold, a recommendation trigger instruction is generated; After receiving the trigger instruction, the recommendation strategy generation module combines the user's real-time behavior context to generate a multi-modal recommendation plan including the main recommended content, auxiliary information, and interaction guidance.

2. The recommendation system based on user behavior analysis according to claim 1, wherein The front-end signal acquisition module includes: a multi-source sensor array, a signal dimensionality reduction unit, and an event trigger logic control unit, where: The multi-source sensor array consists of a touch pressure sensor, a cursor trajectory tracker, and a page scrolling monitor, and is used to collect the original sensor data stream of the page scrolling pause duration, touch click interval, and cursor movement acceleration; The signal dimensionality reduction unit is used to divide the original sensor data stream into multiple data blocks according to a time window, each data block has multiple sampling points, and performs dimensionality reduction feature extraction on the data block through principal component analysis (PCA), retains the principal components with a variance contribution rate greater than 85%, and encodes the dimensionality-reduced feature vectors into a pulse density matrix, and the matrix dimension is compressed to 20% - 30% of the original data stream; The event trigger logic control unit activates data upload when any of the following conditions is detected: The page scrolling pause duration > 300 ms and the acceleration change rate < 0.05 m / s 2 ; The touch click interval is < 150 ms for 3 consecutive times; The curvature radius mutation of the cursor movement trajectory > 15° / ms; The data upload outputs the dimensionality-reduced original sensor data stream to the hierarchical pulse processing module. After three-layer collaborative analysis, structured features are output, and finally, the recommendation strategy generation is driven.

3. The recommendation system based on user behavior analysis according to claim 2, wherein The execution steps of the signal dimensionality reduction unit are: Divide the original sensor data stream into data blocks according to a time window t, and each data block contains 500 - 1000 sampling points; Extract features from the data block through principal component analysis (PCA), and set the data block as matrix Q n×p , where n represents the number of samples in the data block, i.e., the number of sampling points, and p represents the number of features in the data block. For matrix Q n×p , perform standardization processing to obtain That is where Q ij refers to the j-th eigenvalue of the i-th sample, and μ j is the mean of the j-th feature, and σ j is the standard deviation of the j-th feature; Calculate the covariance matrix of the standardized matrix where is the transpose matrix of the standardized matrix X; where The eigenvalues λ1 ≥ λ2 ≥ … ≥ λ are obtained from the covariance matrix S p and the corresponding eigenvectors e1, e2, …, e p , and the eigenvalue λ k reflects the amount of information contained in the k-th principal component, and the eigenvector e k determines the direction of the k-th principal component. The variance contribution rate of the k-th principal component where represents the sum of all eigenvalues, and η k represents the proportion of the k-th principal component in the total amount of information. Retain the first m principal components that satisfy to retain the main information of the data.

4. The recommendation system based on user behavior analysis according to claim 1, wherein The hierarchical pulse processing module includes: an instantaneous pulse layer, a behavior wave layer, a trend flow layer, and a cross-layer collaborative control unit, where: The instantaneous pulse layer is set with a window width of w where 10 ≤ w ≤ 100 ms and a sliding step of s where 1 ≤ s ≤ 10 ms. For the behavior signal x(t) in the pulse density matrix obtained after being processed by the signal dimensionality reduction unit, calculate the first-order difference value Δx of the behavior signal within the window. k = x k - x k-1 , when the difference values of three consecutive windows exceed 2 times the standard deviation of the baseline noise level, i.e., Δx k > 2σ noise , it is marked as an invalid instantaneous pulse event; otherwise, it is a valid instantaneous pulse event. σ noise is the standard deviation of the baseline noise level. The behavior wave layer divides the behavior sequence with 1 minute as the time unit, adopts a sliding window mechanism to dynamically adjust the window range to T and 1 ≤ T ≤ 10 minutes, and calculates the entropy value within each window. where p i represents the occurrence probability of behavior sample i within the window, n is the number of samples in the data block. When the entropy change difference ΔH = |H2 - H1| > 0.5 between the entropies H1 and H2 of two adjacent windows, a behavior pattern migration is triggered and it is confirmed as an abnormal behavior operation event, that is, the user's behavior pattern immediately changes from one behavior pattern to another within adjacent time windows, otherwise it is a normal behavior operation event. The trend flow layer constructs a time-gated graph neural network, retains the key behavior paths through a gating mechanism, and predicts the behavior migration path in the next 2 hours; The cross-layer collaborative control unit, when both the instantaneous pulse layer and the behavior wave layer detect invalid and abnormal events, forces the trend flow layer to start real-time graph structure update and allocates an additional 30% of computing resources.

5. The recommendation system based on user behavior analysis according to claim 4, characterized in that, The trend flow layer constructs a time-gated graph neural network, where nodes represent user behavior categories, and edge weights are jointly determined by behavior transition frequency and temporal correlation. That is, for a behavior transition frequency of f ij , and a temporal correlation of r ij , then the weight w ij of edge (i, j) is w = αf ij + (1 - α)r ij , where α is a weight coefficient and 0 < α < 1. Within the range of 1 - 24 hours, the graph structure is dynamically updated every 30 minutes. The key behavior paths with weight w ij > 0.7 are retained through a gating mechanism. When predicting the behavior migration path for the next 2 hours, 60% of the historical trend weight w history and 40% of the real-time behavior entropy change value ΔH are fused to generate a mixed attention coefficient β = 0.6w history + 0.4ΔH. This attention coefficient is used to adjust the degree of attention to historical trends and real-time behavior changes when predicting the behavior migration path.

6. The recommendation system based on user behavior analysis according to claim 1, wherein, The differential signal amplification module includes: a cross-layer information collection unit, a mutual information gain calculation unit, and a trigger instruction generation unit, where: The cross-layer information collection unit collects the processed signal feature data from the instantaneous pulse layer, behavior wave layer, and trend flow layer. For the instantaneous pulse layer, it collects the signal parameters marked as valid instantaneous pulse events. For the behavior wave layer, it collects the entropy change difference and behavior pattern migration information. For the trend flow layer, it collects the node information, edge weights, and hybrid attention coefficients of the time-gated graph neural network; The mutual information gain calculation unit calculates the mutual information between the signal features of each layer, which is used to analyze the mutual information gain between user behavior features and recommended content, and to judge the feedback degree of different user behaviors to different recommended content; When the mutual information gain values of at least two layers among the three-layer behavior signals exceed the preset threshold T g a recommendation trigger instruction is generated, and the instruction includes the correlation information of the signal features of each layer.

7. The recommendation system based on user behavior analysis according to claim 6, wherein The mutual information gain calculation unit sets the signal feature set of the instantaneous pulse layer as X, the signal feature set of the behavior wave layer as Y, and the signal feature set of the trend flow layer as Z. For X and Y, the mutual information where p(x,y) is the joint occurrence probability of X and Y within the window, and p(x) and p(y) are the marginal probability distributions of X and Y respectively. Similarly, I(X;Z) and I(Y;Z) are obtained. At times t and t + 1 in the time window T, by quantifying the change in mutual information between different layers, the cross-layer mutual information gain G is obtained. That is, for the mutual information gain G between the X and Y layers XY (t,t + 1)=I(X;Y) t+1 -I(X;Y) t , and similarly G XZ and G YZ .

8. The recommendation system based on user behavior analysis according to claim 1, wherein The recommended strategy generator module includes: a context analysis unit, a recommended content screening unit, an interaction guidance design unit, and a multi-modal recommendation scheme generation unit, where: The context analysis unit receives the recommendation trigger instruction sent by the differential signal amplifier module and analyzes it in combination with the user's real-time behavior context information, where the real-time behavior context information includes the page where the user is currently located, browsing history, search records, and the device used; The recommended content screening unit screens out the main recommended content that meets the user's needs from the content library according to the context analysis results. At the same time, it screens out relevant auxiliary information for each main recommended content; The interaction guidance design unit designs the interaction guidance method according to the main recommended content and auxiliary information, and the interaction guidance method includes button prompts, pop-up reminders, and recommended list displays; The multi-modal recommendation scheme generation unit integrates the main recommended content, auxiliary information, and interaction guidance method into a multi-modal recommendation scheme, and the multi-modal recommendation scheme is presented to the user in various forms such as text, graphics, video, and audio.

Images