Adversarial simulation learning method for relieving epidemic deviation in multiple behaviors

Abstract

The invention discloses an adversarial simulation learning method for relieving popularity deviation in multiple behaviors, and solves the problem of popularity deviation under multi-behavior recommendation by adopting a multi-task joint optimization form and combining technologies such as an adversarial network and simulation learning. According to the recommendation method, a post-order behavior module is used as a generator, a preorder behavior module is used as a discriminator, a simulation adversarial learning framework is constructed, the generator can generate tail project embedding simulating preorder behavior characteristics, project embedding of preorder behaviors is enhanced through a gating fusion strategy, the recommendation quality is guaranteed, and meanwhile the recommendation efficiency is improved. And finally, the popularity deviation problem is effectively relieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of recommendation systems, popular bias debiasing in user multi-behavior chains, adversarial learning, imitation learning, etc., and specifically to a kind of adversarial simulation learning method for relieving popular bias in multi-behavior. BACKGROUND

[0002] With the rapid development of society in the information age, a large amount of data is generated in daily life. In order to more efficiently process and use these data, personalized recommendation systems have emerged. Taking e-commerce platforms as an example, personalized recommendation systems use multi-stage algorithm processes to select content that meets user preferences from numerous goods for recommendation, which greatly improves the operational efficiency of the platform and enhances user satisfaction. However, personalized recommendation systems are limited by various factors in practical applications. Still focusing on e-commerce platforms, due to factors such as exposure mechanisms, promotional means, and product quality, the recommendation system is difficult to break free from the constraints of the "80-20 rule", i.e., 20% of popular goods attract about 80% of user interaction behavior. This situation makes modern recommendation systems more inclined to recommend popular goods, which violates the original purpose of meeting user personalized needs, and thus generates a popularity bias problem.

[0003] Among the many popularity bias debiasing recommendation methods, adversarial learning is derived from the idea of generative adversarial networks (GAN), which achieves dynamic optimization through the adversarial training of "generator" and "discriminator" and has become an important technical means for debiasing the popularity bias of recommendation systems. The core logic is: through the adversarial process, the recommendation model is forced to "ignore" bias factors such as product popularity and focus on learning user real preferences, thereby alleviating the problem of over-recommendation of popular goods and suppression of cold but high-quality goods.

[0004] Imitation learning is a learning paradigm that optimizes model decision-making by imitating "expert behavior", the core of which is to let the model learn the optimal strategy from "expert demonstration" rather than directly fitting the rules from biased data. In recommendation systems, the essence of popularity bias is that "biased data" (overexposure of popular goods) misleads the model's learning of user real preferences, while imitation learning introduces "unbiased expert strategy" as a demonstration to guide the recommendation model to break away from dependence on popularity and focus on matching user real needs. SUMMARY

[0005] I. Technical problems solved In practical application scenarios, recommendation systems are often influenced by multiple factors such as exposure mechanism, user reviews, and promotion strategies, and are prone to the "80-20 phenomenon", that is, about 20% of popular items account for nearly 80% of user interaction behaviors. This phenomenon makes the system more inclined to recommend popular items, thereby deviating from the core purpose of "precisely meeting the individual needs of users", and ultimately causing popularity bias problems. Existing popularity debiasing methods mostly focus on single behavior recommendation scenarios, while mainstream single behavior debiasing methods are difficult to adapt to the complex sources and transmission characteristics of popularity bias in multi-behavior recommendation, and ultimately fail in multi-behavior recommendation scenarios. In view of the above shortcomings of traditional debiasing recommendation models, the present invention proposes a popularity debiasing model based on adversarial simulation learning: the model takes the subsequent behavior module as the generator and the previous behavior module as the discriminator, constructs an adversarial simulation learning framework, so that the generator can generate tail item embedding that simulates the features of the previous behavior; then through a gating fusion strategy, the item embedding of the previous behavior is enhanced, and finally the popularity bias is effectively alleviated.

[0006] II. Technical solution Step 1, download a public multi-behavior recommendation system dataset covering multiple user interaction behaviors (such as purchase, add to cart, etc.). Perform preprocessing operations: remove user records with interaction frequency below a preset threshold, and re-encode the remaining user and item entities to lay a foundation for the efficiency and accuracy of subsequent analysis and model training.

[0007] Step 2, import the dataset preprocessed in step 1, use graph convolutional networks to aggregate neighborhood information, and simultaneously learn the user's historical interaction data in different behavior dimensions. This process can capture high-order associations and structural features between users and items, and then generate user and item embedding vectors in different behavior scenarios.

[0008] Step 3, count the number of interactions of items in different behavior scenarios in the graph structure, and set a threshold to divide popular items and long-tail items, and finally obtain the item mask matrix.

[0009] Step 4, construct adversarial simulation learning: use the item embedding vector of the previous behavior in the behavior chain as the input of the discriminator, and use the item embedding vector of the subsequent behavior as the input of the generator; introduce the mask matrix obtained in step 3 to carry out adversarial simulation learning, so that the generator can generate embedding vectors that simulate the features of long-tail items in the previous behavior scenario.

[0010] Step 5, design a gating fusion strategy to fuse and enhance the simulated long-tail item embedding vector obtained in step 4 with the item embedding vector corresponding to the behavior in step 3, and finally obtain the enhanced item embedding vector in the corresponding behavior scenario.

[0011] Step 6: Cascade the user behavior chain, incorporating the initial behavior embedding into the aggregation process of the current behavior. After obtaining the final layer embedding representation of the current behavior, perform comparative learning with the enhanced embedding representation of the corresponding behavior in Step 5. This ensures that while acquiring information about the current behavior, key information from the previous behavior is preserved to the greatest extent possible. Subsequently, the embedding representation of the current behavior after comparative learning is used as the initial embedding for the next behavior.

[0012] Step 7: Based on the multi-task framework built in Steps 2 to 6, calculate the total model loss. Construct a weighted loss function by assigning differentiated weights to the loss of each task, and then use the backpropagation algorithm to solve for the gradients of each parameter, providing direction for subsequent model optimization.

[0013] Step 8: Iteratively update the model parameters based on the gradient information from Step 7 to balance the accuracy and novelty of the recommendations and alleviate the popularity bias problem in multi-behavior scenarios. By exploring the model performance under different normalization coefficient settings, and combining multi-dimensional evaluation indicators such as recall, normalized depreciation cumulative gain (NDCG), and popularity bias sensitivity (Nov), the optimal parameter combination is selected to finally generate a recommendation list that meets the user's personalized needs.

[0014] III. Beneficial Effects Compared with existing technologies, the present invention has the following advantages: 1. Current mainstream bias removal methods are mostly limited to single user behavior scenarios (such as clicks, favorites, etc.), making it difficult to adapt to the multi-behavior linkage user decision-making logic in actual recommendation scenarios (such as the "browse-add to cart-purchase" behavior chain). This invention addresses this limitation by extending the bias removal model to multi-behavior collaborative scenarios. By capturing the preference transmission relationship between different behaviors, it achieves more accurate bias removal modeling, ultimately showing significant improvements in core recommendation metrics such as recall and normalized depreciation cumulative gain (NDCG), better meeting the actual business needs for recommendation performance.

[0015] 2. Existing technologies often suffer from sparse user interaction data for long-tail items (non-popular items), leading to their feature representations being overwritten by popular items during model training, thus exacerbating popularity bias. This invention utilizes an adversarial simulation learning framework, constructing a dynamic adversarial mechanism using subsequent behaviors as the generator and preceding behaviors as the discriminator. The generator can specifically generate long-tail item embeddings that simulate preceding behavior preferences, overcoming the limitation of insufficient interaction data for long-tail items. Simultaneously, this mechanism proactively increases the weight of long-tail item features in model training, avoiding feature learning imbalance between popular and long-tail items, fundamentally reducing the bias of "recommending popular items," and achieving unbiased user preference prediction.

[0016] 3. Existing debiasing methods often suffer from trade-offs: either overemphasizing tail-end item recommendations, leading to decreased recommendation accuracy, or focusing on popular item preference mining while neglecting debiasing. This invention achieves a balance through dual optimization: on the one hand, adversarial learning is used to enhance the feature representations of tail-end items, ensuring they are not diluted by popular item features; on the other hand, a gating fusion strategy is used to dynamically filter effective preference information from preceding behaviors. Ultimately, while maintaining high-quality user and item embedding representations, accurate and unbiased recommendations are synergistically achieved, further improving the model's overall recommendation performance and user experience. Attached Figure Description

[0017] Figure 1 Overall flowchart of the present invention

[0018] Figure 2 Overall model framework diagram of the present invention Detailed Implementation

[0019] To clearly illustrate the purpose, technical solution, and advantages of this invention, the invention will be further described in detail below with reference to specific examples and accompanying drawings.

[0020] This invention implements item recommendation based on user multi-behavioral interaction data. The overall process is shown in Figure 1, and the specific steps are as follows: Step 1: Obtain publicly available multi-behavior recommendation datasets from Beibei, Tmall, and Taobao. These datasets primarily contain historical interaction information for three user behaviors: browsing, adding to cart, and purchasing. Perform preprocessing on the multi-behavior datasets: filter users with interaction frequencies below a set standard, and re-unite users and items in the dataset with unique identifiers. Each preprocessed dataset will generate four TXT files: one recording the user's historical interaction information for the three behaviors (view.txt, cart.txt, buy.txt), and one test set file (test.txt). In the historical interaction information files for the three behaviors, each line contains the user ID and the item ID that the user interacted with under the corresponding behavior dimension. In the test set file, the first field of each line is the user ID, and the second field is the item ID of the user's most recent purchase.

[0021] Step 2: Import the preprocessed multi-behavior data from Step 1 and perform domain aggregation operations using a three-layer lightweight graph convolutional architecture. For each specific behavior subgraph... This process integrates higher-order domain information from its subgraph into the learned representations of users and projects to enhance the efficiency of information transfer between nodes. This domain aggregation process can be represented as:

[0022]

[0023] in and These represent the user set and the project set, respectively. and They are respectively represented as in the first Layered propagation of users and projects In behavior Embedded below. Indicates user In behavior A collection of projects that have been interacted with in the process. Existing behavior China and Project A collection of users who have interacted with the user group. After layer graph convolution, the embeddings obtained in each layer are further combined using the mean operation to form the final user and item representations for each behavior:

[0024]

[0025] Step 3: For each action Statistical items Number of interactions under this behavior:

[0026] in The interaction matrix representing each behavior. Indicates user With the project In behavior If there has been an interaction, the value is 0; otherwise, it is 0. Then, for each project... In behavior Define the mask:

[0027] The masks of all projects were then combined into a matrix. ,so It is a diagonal matrix used to select long-tail items.

[0028] Step 4: Construct adversarial simulation learning. First, construct a generator using MLP mapping. : to act Project Embedding Mapping to behavior The embedding space is used, and the popular mask matrix generated in step 3 is used to generate long-tail item embeddings, thereby generating behavior. The simulated long-tail item embedding representation, i.e. Subsequently, a discriminator was constructed using MLP binary classification. The true probability is output for the input embedding, i.e., a confidence score is given for the true tail embedding. Confidence scores are given for simulated tail embedding. .in and These are the parameter sets for the generator and the discriminator, respectively.

[0029] Adversarial training is then performed, with the goal of minimizing the Jensen-Shannon (JS) divergence between the generated and true distributions. A maximum causal entropy constraint is introduced to enhance the generator's generalization ability, thereby preventing mode collapse. in For the generator, conditional input Given the conditional entropy, the loss of the discriminator is:

[0030] The adversarial loss of the generator is:

[0031] Therefore, in behavior The total loss of the adversarial simulation learning is:

[0032] Step 5: Using the generator trained in Step 4, generate behaviors. Simulated long-tail embedding representation Then, a gating fusion strategy is constructed. First, the gating weights are calculated:

[0033]

[0034] in It is a learnable weight matrix. The parameter vector is a learnable vector whose function is to compress the nonlinearly transformed embedding representation into a scalar score, used to calculate the weights. This score is then fused to obtain the behavior. Enhanced Item Embedding Representation :

[0035] Step 6: Construct a behavior cascade network. First, connect the behaviors from Step 3... go through After layer aggregation, the user embedding representation is obtained. and project embedding representation As an action initial input and Using three layers of lightweight graph convolution to transform behavior The aggregation of preference information yields and .

[0036] To ensure that After aggregation under behavior, the previous behavior The information below is not lost, and the user side utilizes the independent embedding representation from step 3. The project side utilizes the enhanced project embedding representation obtained in step five. , respectively with , Constructing contrastive learning, namely:

[0037]

[0038] Therefore, in behavior The total loss of contrastive learning is:

[0039] Step 7: Use Bayesian Personalized Ranking (BPR) loss as the loss for the main supervised task, optimizing the model parameters by maximizing the probability of users ranking positive sample items. Under behavioral conditions, this loss can be calculated as:

[0040] in, Therefore, the final loss function is obtained:

[0041] in, This represents all trainable parameters across all tasks. and These represent the control contrastive learning task and The hyperparameters for regularization ratios are determined. Finally, the gradient values ​​of each parameter are derived through backpropagation, providing a basis for subsequent parameter optimization.

[0042] Step 8: Based on the gradient information obtained in Step 7, the Adam optimizer is used to iteratively optimize the model parameters. The core objective is to further improve the novelty of the recommendation results while ensuring the accuracy of the recommendations. The model performance evaluation revolves around three key indicators: Recall is used to measure the recommendation system's ability to accurately capture users' potential interests; Normalized Discounted Cumulative Gain (NDCG) focuses on the rationality of the item ranking in the recommendation list and evaluates the degree to which the ranking results match the user's true preferences; Novelty (Nov) is used to quantify the model's ability to suppress the bias of "recommending popular items," which is directly related to the novelty and diversity of the recommendation results. Finally, the optimal combination of model parameters is selected through joint evaluation of multiple indicators to generate a recommendation list for users that is both personalized (fitting user interests) and diverse (avoiding popularity bias).

[0043] It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.

Claims

1. An adversarial simulation learning method to mitigate popularization bias in multi-behavior scenarios, characterized by including: The following steps are required: Step 1: Download a publicly available multi-behavior recommendation system dataset that covers various user interaction behaviors (such as purchasing, adding to cart, etc.). Perform preprocessing: remove user records with interaction frequencies below a preset threshold, and re-encode the remaining user and item entities to lay the foundation for efficient and accurate subsequent analysis and model training. Step 2: Import the preprocessed dataset from Step 1, and use a graph convolutional network to aggregate neighborhood information while simultaneously learning historical user interaction data across different behavioral dimensions. This process captures high-order relationships and structural features between users and items, thereby generating user-item embedding vectors for different behavioral scenarios. Step 3: Count the number of interactions of the project in the graph structure under different behavioral scenarios, and set a threshold. By dividing popular projects into long-tail projects, a project mask matrix is ​​finally obtained. Step 4: Construct adversarial simulation learning: Use the item embedding vector of the previous behavior in the behavior chain as the discriminator input and the item embedding vector of the next behavior as the generator input. The mask matrix obtained in step 3 is introduced to carry out adversarial simulation learning, enabling the generator to generate embedding vectors of long-tail item features in the simulation of the previous behavior scenario. Step 5: Design a gating fusion strategy to fuse and enhance the simulated long-tail item embedding vector obtained in Step 4 with the item embedding vector corresponding to the behavior in Step 3, and finally obtain the enhanced item embedding vector under the corresponding behavior scenario. Step 6: Cascade the user behavior chain, incorporating the initial behavior embedding into the aggregation process of the current behavior. After obtaining the final layer embedding representation of the current behavior, perform comparative learning with the enhanced embedding representation of the corresponding behavior in Step 5. This ensures that while acquiring information about the current behavior, key information from the previous behavior is preserved to the greatest extent possible. Subsequently, the embedding representation of the current behavior after comparative learning is used as the initial embedding for the next behavior. Step 7: Based on the multi-task framework built in Steps 2 to 6, calculate the total model loss. Construct a weighted loss function by assigning differentiated weights to the loss of each task, and then use the backpropagation algorithm to solve for the gradients of each parameter, providing direction for subsequent model optimization; Step 8: Iteratively update the model parameters based on the gradient information from Step 7 to balance the accuracy and novelty of the recommendations and alleviate the popularity bias problem in multi-behavior scenarios. By exploring the model performance under different normalization coefficient settings, and combining multi-dimensional evaluation indicators such as recall, normalized depreciation cumulative gain (NDCG), and popularity bias sensitivity (Nov), the optimal parameter combination is selected to finally generate a recommendation list that meets the user's personalized needs.

2. The adversarial simulation learning method for mitigating popularization bias in multi-behavior scenarios according to claim 1, characterized in that, In step 2, for each specific behavior subgraph ,user and projects The domain aggregation process performed on the subgraph is represented as:

3. The adversarial simulation learning method for mitigating popularization bias in multi-behavior scenarios according to claim 1, characterized in that, In step 3, for each action Statistical items The formula for the number of interactions under this behavior is expressed as:

4. The adversarial simulation learning method for mitigating popularization bias in multiple behaviors according to claim 1, characterized in that, In step 3, in behavior The mask formula below is expressed as:

5. The adversarial simulation learning method for mitigating popularization bias in multi-behavior scenarios according to claim 1, characterized in that, In step 4, the loss function of the discriminator is expressed as:

6. An adversarial simulation learning method for mitigating popularity bias in multi-behavior scenarios according to claim 1, characterized in that, In step 4, the generator's loss function is expressed as:

7. The adversarial simulation learning method for mitigating popularization bias in multi-behavior scenarios according to claim 1, characterized in that, In step 5, the gating fusion strategy formula is expressed as:

8. The adversarial simulation learning method for mitigating prevalence bias in multi-behavior scenarios according to claim 1, characterized in that, In step 7, the final loss function is expressed as:

Images