Method for intelligently pushing commodity display according to user behavior habits

Abstract

The invention relates to a method for intelligently pushing commodity display according to user behavior habits, and belongs to the technical field of artificial intelligence and electronic commerce recommendation systems. Aiming at the problems of low recommendation accuracy and insufficient real-time performance caused by dynamic change of user behaviors in the existing commodity pushing technology, the method comprises the following steps of: acquiring multi-dimensional data such as user browsing tracks, click preferences, purchase records and page staying duration, and constructing a dynamic user portrait in combination with a time sequence analysis and clustering algorithm; fusing real-time behavior feedback by adopting an improved collaborative filtering algorithm, mining potential association between behavior characteristics and commodity attributes through a deep learning model, and establishing an adaptive weight adjustment mechanism; and finally, a personalized commodity sorting strategy is generated based on the current scene and behavior trend prediction of the user, and dynamic optimization of the pushed content is realized. The method can be applied to an e-commerce platform, an advertisement putting system and a mobile application, the recommendation accuracy, the user conversion efficiency and the platform sales volume are remarkably improved, and meanwhile computing resource consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and e-commerce recommendation systems, and particularly relates to an intelligent method for pushing products based on user dynamic behavior analysis, which is applicable to personalized recommendation scenarios of e-commerce platforms, advertising delivery systems, and mobile applications. Background Art

[0002] Existing e-commerce platforms generally use algorithms such as collaborative filtering and content recommendation to implement product pushing, but there are still the following technical defects:

[0003] Insufficient dynamic adaptability: Traditional recommendation systems rely on static historical data and are difficult to capture the dynamic changes of users' interests in real time, resulting in lagging recommendation results. For example, after a user frequently browses a certain type of product in a short period and then their interest shifts, the system still continuously recommends the same type of product.

[0004] Incomplete feature extraction: Existing methods have low utilization rates of users' implicit behavior data (such as page stay duration, scrolling speed), and do not effectively combine time series features, resulting in deviation in user interest modeling.

[0005] Cold start and interest drift problems: For new users or low-active users, due to sparse behavior data, the randomness of recommendation results is strong and the accuracy rate significantly decreases; for long-term users, due to interest migration, historical behavior data becomes invalid.

[0006] Low computational efficiency: Real-time recommendation requires frequent updates of user portraits and product association matrices. Traditional algorithms consume high computational resources and it is difficult to balance computational efficiency and recommendation accuracy.

[0007] Therefore, there is an urgent need for an intelligent pushing method that takes into account dynamic adaptability, computational efficiency, and recommendation accuracy. Summary of the Invention

[0008] The present invention aims to solve the problems of low recommendation accuracy and insufficient real-time performance caused by the dynamic changes of user behavior in the prior art, and provides an intelligent method for pushing products based on multi-dimensional behavior analysis, adaptive weight adjustment, and trend prediction, so as to improve the dynamic response ability and computational efficiency of the recommendation system.

[0009] The method of the present invention includes the following steps:

[0010] A1 Multi-dimensional behavior data collection: Real-time collection of users' browsing trajectories, click preferences, purchase records, and page stay duration data on the e-commerce platform through front-end embedding points and log servers. The data formats include JSON structured logs and the original behavior streams stored in a time series database.

[0011] After data cleaning, structured feature fields are extracted, including user ID, behavior type, timestamp, product category, and context information.

[0012] A2 Dynamic User Portrait Construction: Conduct time series analysis on behavioral data, extract periodic behavior patterns (such as browsing home furnishing products after work every day), and use a sliding window mechanism to divide time intervals. The window size is 24 hours and the step size is 1 hour.

[0013] Calculate the dynamic update value of the interest weight by combining the time decay factor. The formula is:

[0014] Wt = Wt-1·e-λΔt + α·Snew

[0015] Where, Wt is the current interest weight, λ is the decay coefficient, Δt is the time interval, Snew is the new behavior score, and α is the learning rate.

[0016] Use the K-means clustering algorithm to group user behavior patterns and construct a clustered portrait to alleviate the cold start problem.

[0017] A3 Improved Collaborative Filtering Algorithm: Divide user behavior into explicit feedback (purchase, collection) and implicit feedback (browsing duration, click times), and calculate the cosine similarity and Jaccard similarity respectively;

[0018] Generate a corrected user similarity matrix by linearly weighting and fusing the two types of similarities, and screen neighbor users to generate an initial recommendation list.

[0019] A4 Deep Learning Association Mining: Construct a two-tower neural network based on the attention mechanism. The user behavior feature encoding tower uses Transformer to extract multi-level behavior features, and the commodity attribute encoding tower uses CNN to process multi-modal data of commodity titles, categories, and prices;

[0020] Calculate the user-commodity association score by cosine similarity to generate a commodity attribute association matrix.

[0021] A5 Adaptive Weight Adjustment and Ranking Optimization: Load preset weight rules according to the user's current scenario (such as promotional activities, search pages), and dynamically adjust the weight ratios of behavioral features, commodity attributes, and timeliness;

[0022] Use the LSTM model to predict the user's behavior trend. The input includes the recent behavior sequence, the time series data of category popularity, and environmental factors (geographical location, device type), and output the commodity priority score and perform dynamic ranking.

[0023] In view of the problems of low recommendation accuracy and insufficient real-time performance caused by the dynamic changes in user behavior in the existing product push technology, this invention collects multi-dimensional data such as user browsing trajectories, click preferences, purchase records, and page residence time, and constructs a dynamic user profile by combining time series analysis and clustering algorithms. An improved collaborative filtering algorithm is adopted to integrate real-time behavior feedback, and the potential correlation between behavior features and product attributes is mined through a deep learning model to establish an adaptive weight adjustment mechanism. Finally, a personalized product ranking strategy is predicted and generated based on the user's current scenario and behavior trend to achieve dynamic optimization of the pushed content. This method can be applied to e-commerce platforms, advertising delivery systems, and mobile applications, significantly improving the recommendation accuracy, user conversion efficiency, and platform sales, while reducing the consumption of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] 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 following drawings 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.

[0025] Figure 1 is a flowchart of a method for intelligently pushing product displays according to user behavior habits involved in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The following will describe in detail the embodiments of the present invention, and the examples of the embodiments are shown in the drawings. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present invention, but should not be construed as a limitation of the present invention.

[0027] Please refer to Figure 1 , the present invention relates to a method for intelligently pushing product displays according to user behavior habits, including the following steps:

[0028] S1: Collection of multi-dimensional user behavior data

[0029] Data source: Real-time collection of user behavior data through front-end data embedding, log servers, and third-party tools (such as Google Analytics), including:

[0030] Explicit behaviors: Purchase records, favorite products, rating feedback;

[0031] Implicit behaviors: Browsing trajectories, page residence time (in milliseconds), click heat zone distribution, sliding frequency;

[0032]

[0033] Context information: access time period, geographical location, device type (mobile / PC), network status (4G / 5G / Wi-Fi).

[0034] Data storage: The cleaned structured data is stored in a time series database (InfluxDB), which supports high-concurrency writing and real-time querying.

[0035] S2: Build a dynamic user profile and features

[0036] Time series analysis:

[0037] Extract the user behavior timestamps and identify periodic patterns (such as browsing electronic products every Friday night);

[0038] Adopt a sliding window mechanism (window size = 24 hours, step size = 1 hour) to dynamically intercept recent data.

[0039] Interest weight update: Calculate the dynamic weight through an exponential decay model.

[0040] Clustering and grouping: Use the K-means algorithm to group user behaviors (such as "high-consumption active group", "price-sensitive group"), and match the new users with the nearest neighbor group profiles.

[0041] S3: Build a commodity feature vector and a recommendation rule library

[0042] Commodity feature extraction:

[0043] Basic attributes: category, brand, price (normalized to 0-1), sales volume, rating;

[0044] Dynamic attributes: real-time click-through rate, add-to-cart rate, promotion label.

[0045] Recommendation rule library:

[0046] Content rule: Based on commodity attribute matching (such as brand preference);

[0047] Collaborative rule: An improved collaborative filtering algorithm (explicit / implicit feedback fusion, weight β = 0.6);

[0048] Hybrid strategy: Combine offline batch calculation (daily update) and real-time recommendation (second-level response).

[0049] S4: Build a recommendation engine model and output a preliminary recommended commodity set

[0050] Algorithm model:

[0051] Two-tower neural network: Combine the user behavior encoding tower (Transformer) and the commodity attribute encoding tower (CNN) with an attention mechanism;

[0052] Deep sorting model: The DIN (Deep Interest Network) is adopted to capture the evolution of user interests.

[0053] Output: Generate a list of the top 200 candidate products, covering long-tail products and popular products.

[0054] S5: Dynamically adjust the product display according to the recommendation context

[0055] Scene-based weight allocation:

[0056] Home page scene: Increase the weight of new products to 60%;

[0057] Search scene: Strengthen the weight of product attribute matching to 70%;

[0058] Promotion scene: Increase the timeliness weight to 65%, and combine LSTM to predict the interest tendency in the next 1 hour.

[0059] Real-time feedback mechanism:

[0060] If the user continuously skips 3 recommended products, trigger real-time weight reduction (the weight of the relevant category - 20%).

[0061] S6: Multi-terminal adaptive display, display the sorted products to the user in a visual way:

[0062] Mobile terminal: Waterfall flow layout, display 6 - 8 products per screen;

[0063] PC terminal: Grid layout, display 12 - 15 products per screen;

[0064] Weak network environment: Load a lightweight page (only text + thumbnail), with a delay of less than 1 second.

[0065] Interaction optimization:

[0066] Automatically adjust the image resolution according to the screen size;

[0067] Real-time monitor the user's sliding speed and dynamically load subsequent products.

[0068] S7: Obtain user feedback and continuously optimize the model:

[0069] Data collection: Record the user's click-through rate (CTR), conversion rate (CVR), and bounce rate;

[0070] Model update:

[0071] Update the user portrait and product features through daily offline training;

[0072] Real-time incremental training (update the sorting model parameters every 5 minutes).

[0073] Exception handling:

[0074] If the CTR is lower than the threshold (5%) for three consecutive times, trigger the recommendation rule for adjusting the reinforcement learning strategy.

[0075] The above-disclosed is only a preferred embodiment of a method for intelligently pushing product displays according to user behavior habits of the present invention. Of course, the scope of rights of the present invention cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention.

Claims

1. A method for intelligently pushing product displays according to user behavior habits, characterized in that, The following steps are involved: Collect multi-dimensional behavioral data of users on e-commerce platforms, including browsing tracks, click preferences, purchase records, and page dwell time; Performing time series analysis on the multi-dimensional behavioral data and combining it with a clustering algorithm to construct a dynamic user profile, wherein the dynamic user profile includes a user interest weight that decays over time; An improved collaborative filtering algorithm is used to process real-time behavior feedback data. The improved collaborative filtering algorithm optimizes neighbor selection by introducing a user behavior similarity correction term. Use deep learning models to mine the potential correlation between user behavior characteristics and product attributes and generate a product attribute correlation matrix; Based on the dynamic user portrait, product attribute association matrix and current user scenario data, the product recommendation priority score is calculated through an adaptive weight adjustment mechanism; The product recommendation priority scores are dynamically sorted and optimized according to the user behavior trend prediction model, and a personalized product display sequence is generated and pushed to the user terminal.

2. The method according to claim 1, wherein The time series analysis includes the following steps: Extract timestamp information from user behavior data and identify periodic behavior patterns; A sliding window mechanism is used to divide the time interval of behavioral data, and the dynamic update value of user interest weight is calculated in combination with the time decay factor.

3. The method according to claim 1, wherein The implementation of the improved collaborative filtering algorithm includes: Divide user behavior into explicit feedback data and implicit feedback data, and calculate user similarity for each; Generating a modified user similarity matrix by weighted fusion of the explicit feedback similarity and the implicit feedback similarity; The target user's neighbor set is screened based on the user similarity matrix to generate an initial product recommendation list.

4. The method according to claim 1, wherein The deep learning model is a dual-tower neural network based on the attention mechanism, including: User behavior feature encoding tower, used to extract multi-level abstract features of user behavior; Product attribute encoding tower, used to extract multimodal features of product title, category, and price; The correlation strength between the user behavior characteristics and the product attribute characteristics is calculated by cosine similarity.

5. The method according to claim 1, wherein The adaptive weight adjustment mechanism includes: Load the preset weight distribution rules based on the type of the user's current access scenario; Based on the deviation between the user's real-time behavior data and historical portrait, the proportion of behavioral feature weight, product attribute weight and timeliness weight in the product recommendation priority score is dynamically adjusted.

6. The method according to claim 1, characterized in that The user behavior trend prediction model uses a long short-term memory network (LSTM), and the input data includes: The user's behavior sequence in the last N sessions; Time series change data on the popularity of product categories; External environmental factor data, including time period, geographic location, and device type.

7. The method according to claim 1, wherein After generating a personalized product display sequence, it also includes: Monitor users’ click-through rate and conversion rate of pushed products; When the click rate is lower than a preset threshold, the recommendation strategy iterative update module based on reinforcement learning is triggered.

Images