Data enhancement recommendation method for inhibiting popularity deviation in multiple behaviors

Abstract

The invention discloses a data enhancement recommendation method for inhibiting popularity deviation in multiple behaviors, which adopts a multi-task joint optimization form and combines causal inference, data enhancement and other technologies to inhibit the popularity deviation problem under multi-behavior recommendation. According to the recommendation method, a novel popularity noise perception weight matrix is set, and intervention weights under different behaviors are adaptively distributed according to popularity of a project. In addition, according to the method, interaction of a user to a tail item is simulated, expression of long-tail item embedding is enhanced, and user preferences in popular items are mined through multi-source data fusion, so that the problem of popular deviation is effectively restrained while recommendation quality is guaranteed.

Description

Technical Field

[0001] The present invention relates to technical fields such as recommendation systems, popularity bias debiasing, causal inference and data enhancement, and specifically designs a data enhancement recommendation method that suppresses popularity bias in multiple behaviors. Background Art

[0002] With the rapid development of the information age, massive amounts of data are emerging in our daily lives. To efficiently process and utilize this data, personalized recommendation systems have emerged. Taking e-commerce platforms as an example, personalized recommendation systems use a multi-stage algorithmic process to filter through a vast array of products and recommend content that matches user preferences, significantly improving platform operational efficiency and user satisfaction. However, in practice, personalized recommendation systems are constrained by various factors. Still taking e-commerce platforms as an example, due to factors such as exposure mechanisms, user reputation, promotional strategies, and product quality, recommendation systems inevitably suffer from the "80 / 20 rule," where approximately 20% of popular products account for approximately 80% of user interactions. This phenomenon causes the system to favor popular products, deviating from its original purpose of satisfying user needs and resulting in popularity bias.

[0003] Among the many debiased recommendation methods, causal inference is widely used as a tool to reveal deep logical connections between variables and quantify causal effects. In recommendation systems, there are two types of causal relationships: user-level and interaction-level. At the user level, causal relationships drive user decision-making processes; at the interaction level, recommendation strategies significantly influence user-system interactions. Causal inference can abstract and quantify popularity biases, allowing them to be mitigated in subsequent operations.

[0004] Data augmentation techniques are also widely used to debias popularity because they can indirectly transfer prior knowledge used in training sample creation to the model, improving its generalization capabilities and enabling better modeling of user representations. Popularity bias in recommendation systems arises from the fact that long-tail items rarely appear in training samples, resulting in insufficient model training for their vectors. Therefore, data augmentation can be used to fully train long-tail items and improve model performance.

[0005] Furthermore, user behavior encompasses not only the ultimate target behavior but also a series of preceding auxiliary actions, such as page browsing, adding items to a shopping cart, and purchasing items on e-commerce platforms. These behaviors reflect user preferences from different perspectives and can serve as auxiliary behaviors to enhance the model, helping it better represent and predict the user's target behavior (typically, a purchase). In recent years, the exploration of popularity debiasing under multiple behaviors aims to effectively suppress bias caused by excessive focus on popularity while accurately understanding and reflecting user interests, thereby improving the overall performance and accuracy of recommendation systems. Summary of the Invention

[0006] 1. Technical Problems Solved

[0007] Because recommendation systems are influenced by multiple factors in real life, such as exposure mechanisms, user word-of-mouth, and promotional strategies, they often exhibit a "28" phenomenon, whereby approximately 20% of popular products account for approximately 80% of user interactions. This phenomenon causes the system to favor popular products, deviating from its original intention of truly meeting users' personalized needs and resulting in popularity bias. Existing popularity debiasing methods mostly consider single-action recommendation scenarios. However, popular single-action debiasing methods are unable to address the complex sources and transmission characteristics of popularity bias in multi-action recommendations, rendering them ineffective in such scenarios.

[0008] To address the aforementioned shortcomings of traditional debiased recommendation models, this paper proposes a popularity debiasing model that leverages causal inference and data augmentation. This model mitigates training biases in popular items through causal effect analysis. By leveraging the popularity of items in user-item interactions, it adaptively assigns intervention weights conditional on item popularity during multi-behavior modeling. Furthermore, data augmentation techniques are employed to suppress the propagation of popularity bias, ultimately improving the model's ability to predict user preferences and improve recommendation accuracy.

[0009] 2. Technical Solution

[0010] Step 1: Download a publicly available dataset for a multi-action recommendation system (including various user behaviors such as purchase, add to cart, and browse). Preprocess the dataset to filter out users with too few interactions and renumber the users and items in the dataset. Finally, use a leave-one-out evaluation method to generate the corresponding training and test set files. Specifically, the test set contains only the user's last interaction item, while the remaining interaction data is used for training.

[0011] Step 2: Import the preprocessed multi-behavior dataset and use graph convolution for neighborhood aggregation to learn the historical interaction data of each user's behavior. By capturing the high-order connectivity and structural connectivity between users and items, we can generate embedded representations of users and items.

[0012] Step 3: Construct a perturbation vector that satisfies the L2 norm constraint. Align the uniformly distributed random basis perturbation with the sign direction of the item embedding vector learned in step 2 to achieve directional perturbation of specific dimensional features. This is used to optimize the random perturbation mechanism of the algorithm to balance the effectiveness and imperceptibility of the perturbation.

[0013] Step 4: Apply a weight constraint to the noise assigned to each item in step 3, and dynamically adjust the popularity perception weight coefficient based on the item popularity to reduce the impact of noise on popular items and enhance the impact on long-tail items. This ensures that both popular and long-tail items receive equal attention during training. At the same time, perform vector concatenation on the user and item representations to which the adjusted noise vectors have been added, thereby obtaining noise-enhanced user and item representations.

[0014] Step 5: Design a new causal graph, set the popular item nodes in the user-item interaction matrix to zero, retaining only the long-tail item nodes. For each long-tail item, generate a true long-tail item embedding by aggregating the features of the users who interact with it. To make the initially simulated long-tail item embedding closer to the true long-tail item embedding in feature space, calculate the loss value of the distance between the two in Euclidean space as a function, and keep the feature space distance within a small range. At the same time, by integrating the initial item embeddings, true long-tail item embeddings, and simulated long-tail item embeddings from step 2 through multi-source data fusion, the optimized item embeddings can not only retain the user preference characteristics of popular items, but also enhance the expressiveness of tail items, effectively suppressing the spread of popularity bias in multi-behavior recommendations.

[0015] In step 6, the final embeddings of user nodes are compared with the initial embeddings. The final embeddings of item nodes are compared with the optimized item embeddings from step 5. The corresponding contrastive losses are adjusted using a softmax temperature coefficient. The final loss is composed of the contrastive losses of both users and items, capturing information differences at different levels to optimize the model learning process. Furthermore, to extract preference information from multi-behavioral data, a behavioral cascade graph convolutional network approach is used to improve model stability and robustness.

[0016] Step 7: Based on the multi-task framework constructed in Steps 2 to 6, calculate the total loss of the model. By assigning differentiated weights to the losses of each task, a weighted loss function is constructed, and the gradient of each parameter is solved using the backpropagation algorithm to provide direction for subsequent optimization.

[0017] In step 8, the model parameters are iteratively updated based on the gradient information from step 7, aiming to balance recommendation accuracy and novelty and mitigate popularity bias in multi-behavior scenarios. By exploring model performance under different normalization coefficient settings and combining multi-dimensional evaluation metrics such as recall, NDCG (normalized discounted cumulative gain), and Nov (sensitivity to popularity bias), the optimal parameter combination is selected to ultimately generate a recommendation list that meets the user's personalized needs.

[0018] 3. Beneficial Effects

[0019] Compared with the existing technology, the present invention has the following advantages:

[0020] 1. Current mainstream popularity debiasing methods only focus on modeling under single behaviors. In contrast, this invention fully considers the actual situation and expands the popularity debiasing to multi-behavior scenarios, which can significantly improve the final recommendation performance.

[0021] 2. By combining noise enhancement with popularity-aware weighting, compared to existing technologies, the model can fully consider the popularity of items under different user behaviors, adaptively alleviate the complex sources of popularity deviation in multi-behavior scenarios, and comprehensively consider user preferences, thereby providing a foundation for subsequent simulation projects.

[0022] 3. By simulating interactions with existing tail items, the model can enhance the feature representation of tail items and utilize multivariate data fusion strategies to mine hidden user preferences in popular items. This ensures unbiased recommendations while maintaining high-quality user and item embedding representations, ultimately improving the model's recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Overall flow chart of the present invention

[0024] Figure 2 A new causal diagram designed by the present invention

[0025] Figure 3 Overall model framework diagram of the present invention DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to specific examples and accompanying drawings.

[0027] This invention recommends items based on user multi-behavior interaction data. The overall process is as follows: Figure 1 The specific steps include:

[0028] Step 1: Download the publicly available Beibei, Tmall, and Taobao multi-behavior recommendation datasets. These datasets primarily contain historical interaction information for four user behaviors: browsing, adding to cart, and purchasing. Preprocess the multi-behavior dataset to filter out users with too few interactions, and renumber the users and items in the dataset. Finally, a leave-one-out evaluation method is used to generate the corresponding training and test set files. Specifically, the test set contains only the user's last interaction item, and the remaining interaction data is used for training. Each processed dataset generates four txt files: files containing historical user interaction information for three different behaviors (view.txt, cart.txt, and buy.txt) and a test set file (test.txt). In the interaction information files for each of the four different behaviors, each line consists of a user ID and the item ID that the user interacted with during that behavior. In the test set file, the first value on each line represents the user ID, and the second value represents the ID of the user's most recent purchase.

[0029] Step 2: Import the pre-processed multi-behavior data and use the three-layer lightweight graph convolution architecture to perform neighbor aggregation and map specific behavior subgraphs. The high-order neighbors under are injected into the learned representation of users / items to enhance the information transfer between nodes. Neighbor aggregation can be expressed as:

[0030]

[0031] in and They are represented as the improved embeddings of user u and item i propagated at layer l under behavior t, respectively. represents the set of items that user u has interacted with in behavior t, Table 1 shows the set of users who have interacted with item i in behavior t. After l layers of graph convolution, we use mean pooling to further combine the embeddings obtained in each layer to form the final representation of a user (an item):

[0032]

[0033] Step 3: Initialize the perturbation vector.

[0034] Construct a perturbation vector that satisfies the L2 norm constraint and aligns the random basis perturbation that follows the uniform distribution with the sign direction of the item embedding vector learned in step 2. Specifically, it can be expressed as:

[0035]

[0036] Where sign(·) is a sign function, indicating that the direction of the noise vector deviation will be based on To determine. Implement directional perturbations on specific dimensional features to optimize the random perturbation mechanism of the algorithm to balance the effectiveness and imperceptibility of the perturbation.

[0037] Step 4: Enable the perturbation vector to adaptively perceive popularity.

[0038] According to the perturbation vector generated in step 3, a directional perturbation enhancement of the embedding can be achieved, but the perturbation is uncontrollable.

[0039] In order to perturb the project Δ i Modify the factor to a controllable one, reducing its impact on popular items and increasing its impact on long-tail items, to ensure that both popular and long-tail items receive equal attention during training. Dynamically adjust the popularity perception weight coefficient based on item flow, and impose a weight constraint on the noise assigned to each item in step 3.

[0040] This weight w can be expressed as:

[0041]

[0042] Where p(i) represents the popularity of item i, which is usually measured by the number of degrees of the item in the interaction graph. α is a hyperparameter used to adjust the influence of the noise weight.

[0043] Assign this weight to the project disturbance term Δ i , we can get the item perturbation vector with popularity perception, namely:

[0044] Δ′ i =Δ i ×w.

[0045] Subsequently, the user and item representations with the adjusted noise vectors added are concatenated to obtain the noise-enhanced user embeddings. and project embedding Right now:

[0046]

[0047] Step 5: Simulate the interaction of long-tail projects and enhance project embedding through multi-source fusion technology.

[0048] By analyzing and designing a new causal graph, we found that the reason for the popularity bias is the insufficient representation of long-tail items, which leads to a direct causal relationship between the items and the final recommendation. To obtain the real long-tail item interaction, the popular item nodes in the user-item interaction matrix are set to zero. can be expressed as:

[0049]

[0050] in, Represented as the user-item interaction matrix, I 长尾 Represented as a collection of long-tail items.

[0051] Then, for each long-tail item, the real long-tail item embedding is generated by aggregating the user features that interact with it. It can be expressed as:

[0052]

[0053] In order to ensure that the initial simulated long-tail item embedding is closer to the real long-tail item embedding in the feature space This effectively simulates the user's interaction with real long-tail items. The model calculates the distance between the two in Euclidean space as the loss value of the function, and makes the feature space distance within a small value range. The loss function The expression is as follows:

[0054]

[0055] in, It is represented as a simulated long-tail item embedding, and the loss function is used to bring it closer to the real long-tail item embedding. The distance in Euclidean feature space represents the simulated interaction representation of a user on a long-tail item under behavior t. δ controls the minimum distance threshold to prevent the simulation from being too close, resulting in direct copying of the interaction instead of simulation.

[0056] At the same time, we believe that the popularity of popular items is not only influenced by popularity bias, but also hides a certain amount of user preference information. Real long-tail item embedding in step 5 and simulate long-tail item embedding That is, the enhanced project is embedded It can not only preserve the user preference characteristics of popular projects, but also enhance the expressiveness of tail projects. The specific expression is as follows:

[0057] s i =a T tanh(W·ε i +b), i=1,2,3;

[0058]

[0059] Among them, W,a,b are learnable parameters, α i is the normalized attention weight.

[0060] Step 6: Double contrast learning task, extracting user preference information from multi-behavior data to ensure recommendation accuracy.

[0061] In order to ensure the accuracy of recommendations while removing popularity bias, a dual contrast learning task is implemented on the user side and the item side. The final layer embedding of the user node is contrasted with the initial layer embedding, and the final layer embedding of the item node is contrasted with the optimized item embedding in step 5. The corresponding contrast loss is adjusted by the softmax temperature coefficient. The final loss is composed of the contrast loss of the user and the item, capturing the information differences at different levels to optimize the model learning process. User-side contrast learning task under t behavior Comparative learning tasks with project side Specifically, it can be expressed as follows:

[0062]

[0063] Among them, τ represents the softmax temperature coefficient,

[0064] At the same time, to extract preference information from multi-behavior data, a behavioral cascade graph convolutional network method is used to improve model stability and robustness. Therefore, after the aggregation process under t behaviors, the final embedding representation under t behaviors is used as the initial embedding for the next behavior. Specifically, it can be expressed as:

[0065]

[0066] in, and is represented as a trainable weight matrix, σ(·) is represented as the sigmoid function, and LayerNorm(·) is represented as the layer normalization function. and represent the initial embedding of the next action of action t respectively.

[0067] Subsequently, the Bayesian Personalized Ranking (BPR) loss is used as the loss for the main supervision task to optimize the model parameters by maximizing the probability of users' preference ranking of positive sample items. Under behavior t, the loss can be calculated as:

[0068]

[0069] in, Finally, we get the final loss function L of the behavior set T = {1, 2, 3, ..., t}:

[0070]

[0071] Where Θ represents all trainable parameters in all tasks, λ 1 / 2and μ denote the hyperparameters controlling the contrastive learning task and the L2 regularization ratio, respectively.

[0072] Step 7: Calculate the loss function of the model and calculate the gradient through back propagation.

[0073] Based on the task system constructed in steps 2 to 6, the model loss value is calculated. The gradient value of each parameter is derived through the back-propagation mechanism to provide a basis for subsequent parameter optimization.

[0074] Step 8. Using the Adam optimizer, the model parameters are iteratively optimized based on the gradient information obtained in step 7 to improve the novelty of the recommendation results while maintaining the accuracy of the recommendations. The model performance evaluation is completed through the following indicators: Recall: measures the ability of the recommendation system to accurately capture the user's potential interests; NDCG (Normalized Discounted Cumulative Gain): evaluates the rationality of the item ranking in the recommendation list and the fit with the user's preferences; Nov (popularity bias sensitivity, Novelty): quantifies the model's ability to suppress popularity bias. Through the joint evaluation of multiple indicators, the optimal model parameter combination is screened, and finally a personalized and diverse item recommendation list is generated for the user.

[0075] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention. Therefore, the present invention is not limited to the above-mentioned specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. A data enhancement recommendation method to suppress popular bias in multiple behaviors, which features include The steps are as follows: Step 1: Download a publicly available dataset for a multi-action recommendation system (including various user behaviors such as purchase, add to cart, and browse). Preprocess the dataset to filter out users with too few interactions and renumber the users and items in the dataset. Finally, use a leave-one-out evaluation method to generate the corresponding training and test set files. Specifically, the test set contains only the user's last interaction item, while the remaining interaction data is used for training. Step 2: Import the preprocessed multi-behavior dataset and use graph convolution for neighborhood aggregation to learn the historical interaction data of each user's behavior. By capturing the high-order connectivity and structural connectivity between users and items, we can generate embedded representations of users and items. Step 3: Construct a perturbation vector that satisfies the L2 norm constraint. Align the uniformly distributed random basis perturbation with the sign direction of the item embedding vector learned in step 2 to achieve directional perturbation of specific dimensional features. This is used to optimize the random perturbation mechanism of the algorithm to balance the effectiveness and imperceptibility of the perturbation. Step 4: Apply a weight constraint to the noise assigned to each item in step 3, and dynamically adjust the popularity perception weight coefficient based on the item popularity to reduce the impact of noise on popular items and enhance the impact on long-tail items. This ensures that both popular and long-tail items receive equal attention during training. At the same time, perform vector concatenation on the user and item representations to which the adjusted noise vectors have been added, thereby obtaining noise-enhanced user and item representations. Step 5: Design a new causal graph, set the popular item nodes in the user-item interaction matrix to zero, retaining only the long-tail item nodes. For each long-tail item, generate a true long-tail item embedding by aggregating the features of the users who interact with it. To make the initially simulated long-tail item embedding closer to the true long-tail item embedding in feature space, calculate the loss value of the distance between the two in Euclidean space as a function, and keep the feature space distance within a small range. At the same time, by integrating the initial item embeddings, true long-tail item embeddings, and simulated long-tail item embeddings from step 2 through multi-source data fusion, the optimized item embeddings can not only retain the user preference characteristics of popular items, but also enhance the expressiveness of tail items, effectively suppressing the spread of popularity bias in multi-behavior recommendations. In step 6, the final embeddings of user nodes are compared with the initial embeddings. The final embeddings of item nodes are compared with the optimized item embeddings from step 5. The corresponding contrastive losses are adjusted using a softmax temperature coefficient. The final loss is composed of the contrastive losses of both users and items, capturing information differences at different levels to optimize the model learning process. Furthermore, to extract preference information from multi-behavioral data, a behavioral cascade graph convolutional network approach is used to improve model stability and robustness. Step 7: Based on the multi-task framework constructed in Steps 2 to 6, calculate the total loss of the model. By assigning differentiated weights to the losses of each task, a weighted loss function is constructed, and the gradient of each parameter is solved using the backpropagation algorithm to provide direction for subsequent optimization. In step 8, the model parameters are iteratively updated based on the gradient information from step 7, aiming to balance recommendation accuracy and novelty and mitigate popularity bias in multi-behavior scenarios. By exploring model performance under different normalization coefficient settings and combining multi-dimensional evaluation metrics such as recall, NDCG (normalized discounted cumulative gain), and Nov (sensitivity to popularity bias), the optimal parameter combination is selected to ultimately generate a recommendation list that meets the user's personalized needs.

2. The data enhancement recommendation method for suppressing popularity deviations in multiple behaviors according to claim 1 is characterized in that: In step 2, the graph convolution formula performed on the interaction graph between user u and item i is expressed as follows:

3. The data enhancement recommendation method for suppressing popularity deviations in multiple behaviors according to claim 1 is characterized in that: In step 3, the random basis perturbation generation following uniform distribution is expressed as follows:

4. The data enhancement recommendation method for suppressing popularity deviations in multiple behaviors according to claim 1 is characterized in that: In step 4, the assigned noise weight constraint matrix is ​​expressed as follows: Where p(i) represents the popularity of item i, which is usually measured by the number of degrees of the item in the interaction graph. α is a hyperparameter used to adjust the influence of the noise weight.

5. The data enhancement recommendation method for suppressing popularity deviations in multiple behaviors according to claim 1 is characterized in that: In step 5, the real long-tail item interaction matrix It is expressed as follows: in, Represented as the user-item interaction matrix, I 长尾 Represented as a long-tail item collection.

6. The data enhancement recommendation method for suppressing popularity deviations in multiple behaviors according to claim 1 is characterized in that: In step 5, the model calculates the distance between the two in Euclidean space as the loss value of the function, and makes the feature space distance within a small value range. The loss function The expression is as follows: in, It is expressed as simulating long-tail item embedding, and δ controls the minimum distance threshold.

7. The data enhancement recommendation method for suppressing popularity deviations in multiple behaviors according to claim 1 is characterized in that: In step 5, the initial embedding of the project is integrated through multi-source data fusion Real long-tail project embedding and simulate long-tail item embedding The expression is as follows: Among them, W,a,b are learnable parameters, α i is the normalized attention weight.

8. The data enhancement recommendation method for suppressing popularity deviations in multiple behaviors according to claim 1 is characterized in that: In step 6, the final learning objective is the joint optimization of the main recommendation task loss, the simulated tail item loss, and the contrastive learning loss, which is defined as: Where Θ represents all trainable parameters in all tasks, λ 1 / 2 and μ denote the hyperparameters controlling the contrastive learning task and the L2 regularization ratio, respectively.

Images