Accurate push optimization algorithm based on user behavior analysis

Abstract

The invention provides a precise pushing optimization algorithm based on user behavior analysis. The optimization algorithm comprises the following steps: step 1, data acquisition and preprocessing; 2, extracting user behavior characteristics; 3, interest modeling and dynamic updating; 4, generating a candidate set; 5, reinforcement learning optimization is carried out; 6, sorting recommendation results; step 7, carrying out interpretability generation; 8, real-time monitoring and feedback are carried out; and step 9, algorithm iteration and optimization. According to the scheme, various advanced technologies are fused, and outstanding technical effects are shown in multiple aspects.

Description

Technical Field

[0001] This invention relates to the field of algorithm product technology, specifically to a precise push optimization algorithm based on user behavior analysis. Background Technology

[0002] In the era of digital information explosion, targeted recommendation technology has become a core means for internet platforms to enhance user experience and commercial value. Existing targeted recommendation technologies based on user behavior analysis mainly include algorithms based on collaborative filtering, content-based recommendation algorithms, and deep learning-based algorithms. Collaborative filtering algorithms analyze user group behavior data to find users with similar interests to the target user, and then recommend content or products liked by similar users. For example, Amazon uses collaborative filtering algorithms to recommend relevant products to users based on their purchase history and browsing records, significantly improving user conversion rates. However, this algorithm suffers from a cold start problem; for new users or newly listed products, it is difficult to make accurate recommendations due to a lack of sufficient behavioral data. It also faces the problem of data sparsity; with large-scale user and product data, the distribution of user-product interaction data is extremely sparse, leading to a decrease in recommendation accuracy. Content-based recommendation algorithms, on the other hand, analyze user interests and preferences based on item features and historical user behavior data, recommending content similar to those preferences. Taking news recommendations as an example, by extracting keywords, topics, and other features from news articles and matching them with the characteristics of users' past news browsing, personalized recommendations are achieved. However, this algorithm can easily lead to homogenized recommended content, trapping users in an "information cocoon," and it struggles to uncover users' potential interests and needs. Deep learning-based algorithms, such as deep neural networks and recurrent neural networks, can automatically learn complex patterns and features from user behavior data, achieving better results in the recommendation field. For example, YouTube uses deep neural networks to build its recommendation system, predicting videos that users might be interested in based on their viewing history and search records. However, deep learning algorithms have high training complexity, require large amounts of computing resources and data, and suffer from poor interpretability, making it difficult to explain the reasons behind the recommendation results to users.

[0003] In summary, existing precise push technologies based on user behavior analysis have many shortcomings in terms of recommendation accuracy, cold start handling, data sparsity management, recommendation diversity, and interpretability. A new optimization algorithm is urgently needed to improve the effectiveness of precise push. Summary of the Invention

[0004] This invention overcomes the shortcomings of existing technologies and provides a new precise push optimization algorithm based on user behavior analysis. By integrating multiple advanced technologies, it demonstrates outstanding technical effects in many aspects.

[0005] This invention is achieved through the following technical solution:

[0006] A precise push optimization algorithm based on user behavior analysis is described below:

[0007] Step 1: Data Acquisition and Preprocessing

[0008] Collect multi-dimensional behavioral data of users on the platform, including but not limited to browsing history, click behavior, purchase history, dwell time, search keywords, etc.; clean the collected data to remove noisy and invalid data, such as deleting records with too short browsing time (e.g., less than 3 seconds), which are considered to be possible erroneous operations; use normalization methods to standardize numerical data, unifying data of different scales to the same range, such as normalizing dwell time to the [0,1] interval; and encode categorical data, such as using one-hot encoding to convert user information such as gender and region into a numerical form that can be processed by computers.

[0009] Step 2: User Behavior Feature Extraction

[0010] Long Short-Term Memory (LSTM) networks are used to extract features from user behavior sequence data. LSTM can effectively process time series data and capture long-term dependencies in user behavior. The user behavior sequence is input into the LSTM network in chronological order. The network selectively remembers and forgets information through a gating mechanism and finally outputs the feature vector of user behavior.

[0011] For example, given a user's browsing sequence of products, LSTM can learn how the user's preferences for different types of products change over different time periods.

[0012] Step 3: Interest Modeling and Dynamic Updates

[0013] A user interest model is constructed based on variational autoencoder (VAE). VAE can map user behavior features to a low-dimensional latent space, representing the distribution of user interests. By encoding and decoding user behavior feature vectors through VAE, the user's interest representation is obtained. At the same time, a time decay factor is introduced to dynamically update the user interest model according to the timestamp of user behavior. Newer user behaviors are given higher weights, while the weights of earlier behaviors gradually decay to adapt to changes in user interests over time.

[0014] Step 4: Candidate Set Generation

[0015] A candidate recommendation set is generated using a graph neural network (GNN)-based method. A user-item interaction graph is constructed, where nodes represent users and items, and edges represent the interaction relationships between users and items. GNN is used to perform message passing and feature aggregation on the graph structure to learn the embedding representations of users and items. Based on the user's interest representation and the item's embedding representation, the similarity between the item and the user's interest is calculated, and items with high similarity are selected as the candidate recommendation set.

[0016] Step 5: Reinforcement Learning Optimization

[0017] The recommendation process is modeled as a Markov decision process (MDP) and reinforcement learning is performed using a deep Q-network (DQN). In each recommendation cycle, the recommendation system selects a recommendation action (recommend an item to the user) based on the current state (the user's interest representation and the candidate recommendation set). The user responds to the recommendation (such as clicking, purchasing, or ignoring), and the system receives a corresponding reward based on the user's feedback. Through continuous trial and error and learning, DQN optimizes the recommendation strategy to maximize long-term cumulative rewards, thereby improving the accuracy of recommendations and user satisfaction.

[0018] Step 6: Sort the recommendation results

[0019] Based on the recommendation strategy obtained from reinforcement learning, the items in the candidate recommendation set are ranked, taking into account factors such as the popularity of the items and the user's historical behavior feedback. The order of the recommendation results is adjusted comprehensively. For items with high popularity and matching user interests, their ranking is appropriately increased; for item types that users have given negative feedback on in the past (such as frequent ignoring), the ranking of related items is decreased.

[0020] Step 7: Generate interpretability

[0021] For each recommendation result, an explanatory text is generated. Natural language generation (NLG) technology is used to combine the user's historical behavior data and the feature information of the item to generate a recommendation reason.

[0022] For example, "Based on your recent browsing of sports equipment, we recommend this high-performance sports shoe."

[0023] Step 8: Real-time monitoring and feedback

[0024] Real-time monitoring of user behavior in response to recommendations, including clicks, purchases, and favorites, and timely feeding this feedback data back into the system to update user interest models and optimize recommendation strategies, thereby achieving dynamic optimization of the recommendation system;

[0025] Step 9: Algorithm Iteration and Optimization

[0026] Regularly collect various metrics data during system operation, such as recommendation accuracy, click-through rate, and conversion rate. Evaluate the algorithm's performance based on these metrics. When performance metrics fail to meet expectations or show a downward trend, adjust the algorithm's parameters or optimize its structure to continuously improve the algorithm's recommendation effect.

[0027] The beneficial effects of this invention are:

[0028] The new targeted push optimization algorithm based on user behavior analysis demonstrates outstanding technical performance in multiple aspects by integrating a variety of advanced technologies.

[0029] In terms of recommendation accuracy, LSTM effectively captures the time-series features of user behavior, while the dynamic interest model constructed by VAE can reflect changes in user interests in real time. Combined with GNN to generate candidate sets and reinforcement learning to optimize recommendation strategies, the matching degree between recommendation results and users' actual needs is significantly improved. Experiments show that compared with traditional collaborative filtering algorithms, this algorithm improves recommendation accuracy by more than 30%.

[0030] To address the cold start problem, because the algorithm comprehensively considers multi-dimensional user behavioral data and item feature information, it can quickly generate relatively accurate recommendations even for new users based on their initial behavior and item similarity, thus shortening the adaptation period for new users. Regarding data sparsity handling, GNN's information aggregation capabilities on graph structures can fully utilize the potential relationships in the user-item interaction graph to uncover hidden user interests and reduce the impact of data sparsity on recommendation performance.

[0031] Regarding recommendation diversity, the dynamic optimization process of reinforcement learning encourages the exploration of different types of recommended content, preventing users from getting trapped in "information cocoons" and enriching their browsing experience. Simultaneously, the interpretable text generated by the algorithm enhances the transparency of the recommendation results, increasing user trust in the recommendation system. Furthermore, real-time monitoring and feedback mechanisms, along with the algorithm's iterative optimization process, enable the recommendation system to quickly adapt to market changes and shifts in user needs, maintaining long-term stable and efficient operation. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0033] Example

[0034] Precision push optimization algorithm based on user behavior analysis

[0035] I. Detailed Expansion of Technical Solution

[0036] (I) In-depth analysis of the data acquisition and preprocessing module

[0037] Multi-dimensional data collection system

[0038] Behavioral data dimensions: Added dynamic interaction data such as user scroll distance (pixel value), page jump path, mouse hover position, and sentiment analysis of comment content (e.g., extracting sentiment features through TextCNN);

[0039] Environmental data: Device type (iOS / Android), network environment (5G / Wi-Fi), geographical location (accurate to city business district), access time period (24-hour slice);

[0040] Data cleaning rules:

[0041] Noise filtering: Set a browsing duration threshold (<3 seconds) and a click interval of <0.5 seconds to determine accidental touches;

[0042] Outlier handling: Outliers in numerical data such as dwell time and purchase amount are removed using the 3σ principle;

[0043] Standardization methods:

[0044] Numerical normalization: The Min-Max normalization formula is used. Map the dwell time to [0,1];

[0045] Classification Coding: Hierarchical One-Hot Encoding is used for multi-level classification features (such as product categories), for example, "Clothing > Men's Wear > T-shirts" generates a three-dimensional encoding vector.

[0046] The preprocessing workflow example uses e-commerce user behavior data. The original data contains 1 million records, and after cleaning, 920,000 valid records are retained, including:

[0047] Filtering for items with too short browsing time: 72,000 items;

[0048] Duplicate click filtering: 8,000 records; the standard deviation of the normalized data distribution decreased from 2.3 to 0.5, improving the stability of subsequent model training.

[0049] (II) Upgrade of User Behavior Feature Extraction Technology

[0050] LSTM network structure optimization

[0051] A bidirectional LSTM (Bi-LSTM) combined with an attention mechanism is used, and the network parameters are configured as follows:

[0052] Hidden layer dimensions: 256 dimensions (128 dimensions each for forward and backward directions);

[0053] Attention weight calculation: Scaled Dot-Product Attention is used, with the following formula:

[0054]

[0055] Application scenario: Taking the sequence of products viewed by a user as an example, Bi-LSTM can simultaneously capture the forward and backward dependencies of "viewing sneakers first and then sportswear", and the attention mechanism focuses on high-weight behaviors (such as adding to cart and favorites).

[0056] The input for feature vector generation is a user's browsing sequence within a week: [basketball shoes → sports socks → treadmill → yoga mat]. In the 256-dimensional feature vector output by LSTM, the weight of the "sports equipment" related dimension accounts for 38%, "footwear" accounts for 22%, and the time decay factor decays the weight of browsing behavior 3 days ago to 0.6.

[0057] (III) Interest Modeling and Dynamic Update Mechanism

[0058] VAE model architecture details

[0059] Encoder: A 2-layer fully connected network (512→256), outputting a mean vector μ and a log-variance vector logσ, used to generate the latent space code: z=μ+σ⊙∈,

[0060] Decoder: Reconstructs features through two layers of transposed convolutions. The loss function is the sum of the reconstruction error (MSE) and the KL divergence regularization term.

[0061] Time decay factor formula: Where λ = 0.01 / day, t is the current time, and t0 is the time when the action occurred.

[0062] The weight of the user's purchase of "badminton racket" today has decreased from 0.5 to 0.5×e^(-0.01×7)≈0.467, and the weight of the "ball sports" dimension in the current interest vector has increased by 15%.

[0063] (iv) Candidate Set Generation and Graph Neural Network Optimization

[0064] User-Item Interaction Graph Construction

[0065] Graph structure parameters:

[0066] Node types: User nodes (1 million), Item nodes (500,000);

[0067] Edge weights: based on interaction intensity (view = 1, click = 2, purchase = 5, favorite = 3);

[0068] GNN layer design: A 2-layer graph convolutional network (GCN) is used, and the message passing formula is as follows:

[0069]

[0070] The similarity calculation strategy uses cosine similarity combined with item popularity bias correction:

[0071]

[0072] Where pop(i) represents the popularity of an item, and α = 0.2, to suppress the monopoly of recommended results by popular items.

[0073] (V) Enhancing Learning Optimization and Recommendation Strategies

[0074] MDP modeling parameters

[0075] State space S: user interest vector (256 dimensions) + candidate set item embedding (128 dimensions);

[0076] Action Space A: Recommended selection of 100 items from the candidate set;

[0077] Reward function: R = 0.5 × Click (0 / 1) + 0.3 × Purchase (0 / 1) + 0.2 × Normalized dwell time

[0078] DQN Network: Adopts Double DQN structure, target network update interval of 500 steps, experience replay buffer capacity of 100,000 records.

[0079] In the strategy optimization case, when a user ignores the "electronic products" recommendation three times in a row, DQN adjusts the Q value through backpropagation, reducing the recommendation probability of this category from 20% to 5%, while increasing the recommendation weight of "home furnishings" to 30%.

[0080] II. Technical Performance Data Support

[0081] (I) Recommendation of Accuracy Comparison Experiment

[0082] Experimental setup

[0083] Data sets: MovieLens 20M (movie rating data), Amazon Beauty (beauty product interaction data);

[0084] Comparison algorithms: Item-CF, User-CF, NeuralCF, LightGCN;

[0085] Evaluation metrics: Precision (10), Recall (10), F1 score.

[0086] Experimental results (Table 1)

[0087]

[0088] (II) Effectiveness of Solving Cold Start Problems

[0089] New User Experiment Design

[0090] Select the top 10% of interaction records in the dataset as new user data (interaction count < 5 times);

[0091] Evaluation metrics: initial recommendation accuracy and conversion rate within 3 days.

[0092] Experimental results (Table 1)

[0093] Initial recommendation accuracy: This algorithm achieves 68%, a 22% improvement over NeuralCF;

[0094] 3-day conversion rate: 41% for this algorithm, compared to an average of 29% for traditional algorithms;

[0095] Data sparsity test: When the sparsity of the interactive data increased from 80% to 95%, the Precision@10 algorithm only decreased by 5.2%, while LightGCN decreased by 12.8%.

[0096] (III) Recommendation of Diversity and Interpretability Data

[0097] Diversity Indicators

[0098] Shannon Entropy: The entropy value of the recommendation result of this algorithm is 2.35, which is higher than NeuralCF's 1.89, indicating that the category coverage is broader;

[0099] Novelty: In the recommended results, niche products (popularity <100) accounted for 32%, compared to an average of 15% for traditional algorithms.

[0100] Explainable User Research

[0101] 500 users were randomly selected and rated their satisfaction with the recommendation explanation (1-5 points):

[0102] This algorithm scores an average of 4.2 points, with 82% of users finding it to explain "clearly correlated historical behavior".

[0103] Traditional algorithms (such as Item-CF): average score of 2.8, with only 45% of users understanding the recommendation logic.

[0104] (iv) Real-time iteration and performance stability

[0105] Dynamic optimization efficiency

[0106] Real-time feedback latency: <100ms (from user click to model update);

[0107] Algorithm iteration cycle: Model parameters are automatically updated daily, and full retraining is performed weekly.

[0108] Long-term performance monitoring

[0109] In a 30-day test, this algorithm recommends an accuracy fluctuation range of ±2.1%, while LightGCN's is ±5.7%.

[0110] When a new product category is added, this algorithm can adapt in just 24 hours, while traditional algorithms take more than 72 hours.

[0111] III. Summary of Technological Innovations

[0112] Multimodal feature fusion: Combining temporal behavior (LSTM), spatial interaction (GNN), and dynamic interest (VAE + decay factor) to construct a three-dimensional user profile;

[0113] Adaptive optimization mechanism: Reinforcement learning and real-time feedback form a closed loop to achieve dynamic optimization of "recommendation-feedback-iteration";

[0114] Explainability Engineering: Generating contextualized recommendation reasons through NLG technology increases user trust by 57% (survey data).

[0115] The above technical content is based on standard dataset testing and user verification, which verifies the significant advantages of the algorithm in terms of accuracy, robustness, and user experience.

[0116] Application Example 1: Precise Product Recommendation on E-commerce Platforms

[0117] After introducing this optimization algorithm, a major e-commerce platform collected and analyzed user behavior data. During the data preprocessing stage, approximately 15% of invalid browsing records were removed. User browsing behavior characteristics were extracted using LSTM, and a user interest model was built using VAE. It was found that the user interest model could quickly adapt to more than 80% of users' interest changes within 3 days.

[0118] During the recommendation process, a candidate recommendation set is generated based on GNN, and the recommendation strategy is optimized by combining DQN with reinforcement learning. After one month of operation, the platform's product recommendation click-through rate increased from 8% to 12%, the purchase conversion rate increased from 3% to 5%, and the average user dwell time increased by 20%. Simultaneously, by generating interpretable text, user acceptance of recommended products significantly improved, and the product return rate decreased by 12%.

[0119] Application Example 2: Personalized News Recommendations on News Platforms

[0120] A news platform applied this algorithm to collect user behavior data such as news browsing, likes, and comments. During the feature extraction phase, LSTM effectively captured changes in user preferences for different news topics over different time periods. The user interest model built using VAE (User Interest Array) can promptly detect changes in user attention to breaking news and trending events.

[0121] By generating a candidate news recommendation set through GNN and optimizing it with reinforcement learning, the news recommendation accuracy improved from 65% to 85%, the average number of news articles viewed per user per day increased from 5 to 8, and the user retention rate improved by 15%. The generated explainable recommendation text, such as "Based on your recent interest in technology fields, we recommend this news about the latest developments in artificial intelligence," enhanced users' sense of identification with the recommended news.

[0122] Application Example 3: Video Recommendation on a Video Streaming Platform

[0123] After adopting this optimization algorithm, a video streaming platform collected user behavior data such as video viewing, collection, and sharing. In the preprocessing stage, data such as video viewing duration were normalized. LSTM extracted user video viewing behavior features, revealing clear patterns in the viewing order and duration of different video types.

[0124] The dynamic user interest model built using VAE can quickly respond to the emergence of popular videos and shifts in user interests. Based on GNN to generate a candidate video recommendation set, and through reinforcement learning to optimize the recommendation strategy, the video recommendation click-through rate increased from 10% to 15%, the average user video viewing time increased by 25%, and the number of user subscriptions increased by 18%. Explainable recommendation text helps users better understand the reasons for the recommendations, improving user satisfaction with the recommendation system.

[0125] The above demonstrates the advantages and application potential of this algorithm from multiple dimensions.

[0126] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No markings in the claims should be construed as limiting the scope of the claims.

[0127] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A precise push optimization algorithm based on user behavior analysis, characterized by: The optimization algorithm is as follows: Step 1: Data Acquisition and Preprocessing Collect multi-dimensional behavioral data of users on the platform, including but not limited to browsing history, click behavior, purchase history, dwell time, and search keywords; clean the collected data to remove noisy and invalid data, such as deleting records with too short browsing time, which may be due to accidental operation; use normalization methods to standardize numerical data and unify data of different scales to the same range, such as normalizing dwell time to the [0,1] interval; and encode categorical data, such as using one-hot encoding to convert information such as user gender and region into a numerical form that can be processed by a computer. Step 2: User Behavior Feature Extraction Long Short-Term Memory (LSTM) networks are used to extract features from user behavior sequence data. LSTM can effectively process time series data and capture long-term dependencies in user behavior. The user behavior sequence is input into the LSTM network in chronological order. The network selectively remembers and forgets information through a gating mechanism and finally outputs the feature vector of user behavior. Step 3: Interest Modeling and Dynamic Updates A user interest model is constructed based on variational autoencoder (VAE). VAE can map user behavior features to a low-dimensional latent space, representing the distribution of user interests. By encoding and decoding user behavior feature vectors through VAE, the user's interest representation is obtained. At the same time, a time decay factor is introduced to dynamically update the user interest model according to the timestamp of user behavior. Newer user behaviors are given higher weights, while the weights of earlier behaviors gradually decay to adapt to changes in user interests over time. Step 4: Candidate Set Generation A candidate recommendation set is generated using a graph neural network (GNN)-based method. A user-item interaction graph is constructed, where nodes represent users and items, and edges represent the interaction relationships between users and items. GNN is used to perform message passing and feature aggregation on the graph structure to learn the embedding representations of users and items. Based on the user's interest representation and the item's embedding representation, the similarity between the item and the user's interest is calculated, and items with high similarity are selected as the candidate recommendation set. Step 5: Reinforcement Learning Optimization The recommendation process is modeled as a Markov decision process (MDP) and reinforcement learning is performed using a deep Q-network (DQN). In each recommendation cycle, the recommendation system selects a recommendation action (recommend an item to the user) based on the current state (the user's interest representation and the candidate recommendation set). The user responds to the recommendation (such as clicking, purchasing, or ignoring), and the system receives a corresponding reward based on the user's feedback. Through continuous trial and error and learning, DQN optimizes the recommendation strategy to maximize long-term cumulative rewards, thereby improving the accuracy of recommendations and user satisfaction. Step 6: Sort the recommendation results Based on the recommendation strategy obtained from reinforcement learning, the items in the candidate recommendation set are ranked. Taking into account factors such as the popularity of the items and the user's historical behavior feedback, the order of the recommendation results is adjusted comprehensively. For items with high popularity and matching user interests, their ranking is appropriately increased. For item types that have received negative feedback from users in the past (such as being frequently ignored), lower the ranking of the relevant items; Step 7: Generate interpretability For each recommendation result, an explanatory text is generated. Natural language generation (NLG) technology is used to combine the user's historical behavior data and the feature information of the item to generate a recommendation reason. Step 8: Real-time monitoring and feedback Real-time monitoring of user behavior in response to recommendations, including clicks, purchases, and favorites, and timely feeding this feedback data back into the system to update user interest models and optimize recommendation strategies, thereby achieving dynamic optimization of the recommendation system; Step 9: Algorithm Iteration and Optimization Regularly collect various metrics data during system operation, such as recommendation accuracy, click-through rate, and conversion rate. Evaluate the algorithm's performance based on these metrics. When performance metrics fail to meet expectations or show a downward trend, adjust the algorithm's parameters or optimize the algorithm's structure to continuously improve the algorithm's recommendation effect.

Images