A popularity debiasing recommendation model based on item similarity

Abstract

The application provides a popularity debiased recommendation model based on item similarity, which comprises the following steps: extracting user item rating data from a database and preprocessing the data to obtain an original rating matrix; calculating the similarity between items based on KL divergence, and calculating the popularity based on item similarity (ISP) based on the similarity; based on the causal theory, decoupling the interaction score of the user and the item into two parts of "real interest" and "herd behavior", and assigning corresponding embedding vectors to the user and the item; by introducing a weight attention mechanism, adaptively learning the weight of the user's interest and herd behavior; based on the obtained weight, weighted sum of the two parts of the score, calculate the final personalized recommendation score; by using the negative sampling method based on ISP, the training data set is divided, and the loss function is constructed to supervise the learning of the decoupled embedding vector; predicting and ranking the score of the item that the user has not rated, and selecting the top k items to form a personalized recommendation list.

Description

TECHNICAL FIELD

[0001] The present application relates to the recommendation technical field, and particularly relates to a popularity debiasing recommendation model based on item similarity. BACKGROUND

[0002] A recommendation system plays an important role in solving information overload, improving user experience, increasing user stickiness, and becomes one of the indispensable important technologies in the information age. The existing recommendation model usually models user interest based on observed user-item interaction data, and the observed data may also be due to the conformity behavior of users to popular items. Therefore, when learning user interest, the traditional recommendation model is often affected by item popularity, so it is difficult to accurately capture the real interest of users, resulting in reduced accuracy of recommendation and causing popularity bias problem.

[0003] To solve the popularity bias problem, some existing debiasing methods reduce bias through data preprocessing, regularization, reweighting training samples and other techniques. The existing methods still have the following problems when dealing with the decoupling of interest and conformity behavior: first, in the observed data, the interest and conformity behavior of users are usually mixed together, so it is important to find a reasonable signal to distinguish the two for effective training on the original observed data. However, these methods are based on traditional global item popularity, and only use interaction frequency to calculate popularity, without considering the actual popularity of items from a more fine-grained level. This may lead to inaccurate distinction between interest and conformity behavior, thereby affecting the training effect of the model. Second, the existing debiasing recommendation model models user-item interaction behavior as a simple addition of interest score and conformity behavior score, which is difficult to represent the personalized behavior of users, affects the expression ability of the model, and further affects the recommendation performance.

[0004] Therefore, a popularity debiasing recommendation model based on item similarity is designed, which generates more reasonable negative sampling signals by introducing a popularity calculation method based on item similarity, and introduces a weight attention mechanism to learn the attention score of user real interest and conformity behavior, which can adaptively capture the weight of real interest and conformity behavior of specific users, and has great significance for improving the popularity debiasing ability of the recommendation model. SUMMARY

[0005] To solve the above problems, the present application proposes a popularity debiasing recommendation model based on item similarity, which generates more reasonable negative sampling signals based on item similarity, thereby effectively distinguishing the interest and conformity behavior of users in the training process. In addition, an adaptive weight attention mechanism is also designed to capture the personalized behavior of users on interest and conformity behavior. Specifically, it includes the following:

[0006] A data collection and preparation module is configured to acquire and clean the collected data in the database to obtain a required original score matrix.

[0007] A project similarity and popularity calculation module is configured to calculate the project similarity and the project popularity, calculate the project similarity based on the project fine-grained attributes, calculate the project similarity by using a Kullback-Leibler (KL) divergence, dynamically define the "same type" project based on the project similarity, and calculate the relative popularity of the project in the category to which the project belongs, defined as ISP, to obtain a ranking representing the real popularity of the project.

[0008] A causal embedding and decoupling module is configured to learn the interest and the herd behavior of a user by using a causal embedding technology, and decouple the user project interaction behavior into the real interest and the herd behavior.

[0009] A personalized real interest and herd behavior weight calculation module is configured to design an adaptive weight attention mechanism for capturing the personalized real interest and the herd behavior weight of the user.

[0010] A personalized recommendation score prediction module is configured to generate a final prediction value of the user for the item by combining the real interest score and the herd psychology score.

[0011] A popularity-driven model optimization module is configured to train the model parameters, including the construction of a loss function and a negative sampling method. The module provides a supervision signal for learning of the real interest and the herd psychology embedding vector by introducing the popularity information.

[0012] A recommendation result acquisition module is configured to sort the comprehensive prediction value of the user for the un-scored item, select the top k items to generate a recommended item list of the user.

[0013] The project similarity and popularity calculation module considers the fine-grained project attributes, and is configured to calculate the project popularity and the project popularity, and the specific steps are as follows:

[0014] 1) calculate the similarity sim(i,j) between the projects by using the KL divergence;

[0015] 2) select the N projects with the highest similarity to the target project i as the same type project set S i .

[0016] 3) calculate the relative popularity of the target project i in S i , and the calculation formula is as follows:

[0017]

[0018] pop i is the global project popularity of the project i under the traditional definition, and is defined as the interaction frequency in all users.

[0019] The causal embedding and decoupling module described in this invention, from a causal perspective, decouples user-item interaction behavior into genuine interest and conformity behavior, modeling it as a combination of genuine interest scores and conformity behavior scores. It also assigns embedding vectors representing interest and conformity behavior to users and items respectively, helping the recommendation model learn more interpretable user and item features. As shown in the following formula:

[0020]

[0021] Where <·,·> denotes the dot product; Let r be the embedding vectors of genuine interest and conformity, respectively; ui Rate item i for user u; Score for a specific reason.

[0022] The personalized genuine interest and conformity behavior weight calculation module of this invention introduces a multilayer perceptron (MLP) to enhance the model's decoupling capability. This enables the model to capture complex feature relationships and patterns, thereby adaptively learning the optimal weight distribution to accurately reflect the user's genuine interests and conformity behavior. The specific implementation is as follows:

[0023] 1) Given a reason c belonging to a set consisting of users' genuine interests and conformity behaviors, denoted as c∈{int,con}. For a specific reason c, its corresponding embedding... The attention score of this reason is used as input to the MLP. Represented as:

[0024]

[0025] in, All of these are model parameters, where s represents the number of neurons in the hidden layer.

[0026] 2) The attention score is normalized using the softmax function, and the calculation formula is as follows:

[0027]

[0028] The personalized recommendation score prediction module of this invention predicts a user's personalized rating for a project by combining the embedding vectors of user interests and conformity behavior. First, from a causal embedding perspective, the interaction between the user and the project is modeled as a weighted combination of interest scores and conformity behavior scores. Then, the learned embedding vectors are used to calculate the predicted rating for each project. The calculation formula is as follows:

[0029]

[0030] The popularity-driven model optimization module comprises the following parts:

[0031] 1) Loss function construction

[0032] The BPR loss function is used for model training. The training set is composed of triplets (u, i, j), wherein i and j are respectively the items that have been interacted and not interacted by the user u. According to the popularity of the positive and negative samples i and j, the training set can be divided into two cases:

[0033] Case 1: If the popularity of the positive sample i is greater than that of the negative sample j, the corresponding training data set is denoted as D1. It can be considered that the comprehensive score of the positive sample is higher than that of the negative sample The conformity score of the positive sample is higher than that of the negative sample That is, the following inequality holds:

[0034]

[0035] Case 2: If the popularity of the positive sample i is less than that of the negative sample j, the corresponding training data set is denoted as D2. Since the comprehensive score of the positive sample is higher than that of the negative sample The conformity score of the positive sample is lower than that of the negative sample It can be deduced that the interest score of the positive sample is higher than that of the negative sample That is, the following inequality holds:

[0036]

[0037] Based on the above inequality, the following comprehensive loss function is designed to optimize model training.

[0038]

[0039] 2) Item popularity-driven negative sampling

[0040] The item similarity-based popularity ISP calculated by the above module i is used as the basis for negative sampling, and sampling is performed to construct training data sets D1 and D2, so that each training data set can provide effective supervision signals to support the model to learn causal embedding.

[0041] The recommendation result acquisition module sorts the comprehensive scores of the user for the unrated items, and selects the top k items to generate a list of recommended items for the user.

[0042] The application provides a popularity debiased recommendation model based on item similarity, which decouples the interaction behavior of users and items into real interest and herd behavior from the perspective of causality, introduces an adaptive weight attention mechanism, and dynamically adjusts the weight of real interest and herd behavior according to the personalized behavior of users. At the same time, in order to ensure the optimization of the model, the application constructs a reasonable loss function, and calculates the popularity of the item based on the KL divergence, so as to provide a more reasonable supervision signal for the training of the model, guarantee the learning process of the user interest and herd behavior embedding, and further improve the accuracy and debiasing ability of the recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a model process schematic diagram of the application;

[0044] Figure 2 is a model framework diagram of the application;

[0045] Figure 3 is a comparison of HR@5, NDCG@5, HR@10, NDCG@10, PRU@10 of the application with other methods under three data sets. DETAILED DESCRIPTION

[0046] In order to make the purposes, technical solutions and advantages of the application clearer, the application of a neural collaborative filtering recommendation model considering multi-dimensional user preferences is described in detail below in combination with the drawings. It should be understood that the specific implementation method described herein is only used to explain the application, and is not used to limit the application, and the changes, modifications, additions or replacements made by those skilled in the art within the scope of the application should be covered in the scope of the claims of the application.

[0047] Figure 1 is a process schematic diagram of a popularity debiased recommendation model based on item similarity of the application. First, user item rating data is extracted from the database and preprocessed to obtain an original rating matrix. Then, the similarity between items is calculated based on the KL divergence, and the popularity based on item similarity (ISP) is calculated based on the similarity. At the same time, based on the causality theory, the interaction score of users and items is decoupled into real interest and herd behavior, and the corresponding embedding vectors are allocated to users and items. On this basis, by introducing a weighted attention mechanism, the weights of user interest and herd behavior are adaptively learned. Next, the final personalized recommendation score is calculated by weighted sum of the two parts of the obtained weight. The training data set is divided by the negative sampling method based on ISP, and the loss function is constructed for model optimization. After the model training is completed, the rating value prediction and ranking of the items not rated by the user are performed, and the top k items are selected to form a personalized recommendation list.

[0048] Figure 2 is a framework diagram of a popularity debiased recommendation model based on item similarity of the present application. It comprises a data acquisition and preparation module, an item similarity and popularity calculation module, a causal embedding and decoupling module, a personalized real interest and conformity behavior weight calculation module, a personalized recommendation score prediction module, a popularity driven model optimization module, and a recommendation result acquisition module. The data acquisition and preparation module is used to extract data from the database, and perform cleaning and preprocessing operations on the data, etc., to obtain the required original user-item rating matrix; the item similarity and popularity calculation module is used to calculate the KL similarity and popularity between items, to ensure that the popularity accurately reflects the popularity of the items; the causal embedding and decoupling module, based on the causal theory, decouples the interaction scores of users and items into "real interest" and "conformity behavior" two parts, and assigns corresponding embedding vectors to each user and item; the personalized real interest and conformity behavior weight calculation module introduces a weight attention mechanism, and uses a multi-layer perception machine to calculate the attention weights of each user's interest preference and conformity behavior; the personalized recommendation score prediction module, based on the personalized interest and conformity behavior weights obtained in the previous step, generates the final personalized recommendation score, providing a basis for subsequent recommendation list sorting; the recommendation result acquisition module is used to obtain the recommendation list of the target user u. The comprehensive prediction value of the user's un-rated items is sorted, and the top k items are selected to form the personalized recommendation list of the target user u.

[0049] Further, the following cases are taken as examples for illustration:

[0050] The MovieLens dataset has M users U = {u1, u2,..., uM} and N items I = {i1, i2,..., iN}, the rating domain is {1, 2, 3, 4, 5}, and the user-item rating matrix is represented by [R]M x N. M} and N items I = {i1, i2,..., iN}, the rating domain is {1, 2, 3, 4, 5}, and the user-item rating matrix is represented by [R]M x N. N} and N items I = {i1, i2,..., iN}, the rating domain is {1, 2, 3, 4, 5}, and the user-item rating matrix is represented by [R]M x N. M×N} and N items I = {i1, i2,..., iN}, the rating domain is {1, 2, 3, 4, 5}, and the user-item rating matrix is represented by [R]M x N. ui} and N items I = {i1, i2,..., iN}, the rating domain is {1, 2, 3, 4, 5}, and the user-item rating matrix is represented by [R]M x N.

[0051] First, the data acquisition and preparation module is used to obtain and clean the system data to obtain the variables in the hypothesis. The specific implementation steps of the proposed popularity debiased recommendation model based on item similarity are as follows:

[0052] S1: The data acquisition and preparation module obtains the required data information from the database, including user ID, item ID and rating information.

[0053] S2: The item similarity and popularity calculation module is used to calculate the item popularity and item popularity, and the specific steps are as follows:

[0054] 1) Calculate the similarity between items sim(i,j) by KL divergence;

[0055] 2) Define the "similar" item set S of item i i , select the top N items with the highest similarity to item i as its similar item set S based on the calculated similarity i .

[0056] 3) Calculate the relative popularity of target item i in S i , the calculation formula is as follows:

[0057]

[0058] Where pop i is the global item popularity of item i under the traditional definition, defined as the interaction frequency in all users.

[0059] S3: Causal embedding and decoupling module based on causal theory, decouples user item score into real interest score and conformity behavior score , and assigns embedding vectors representing real interest and conformity behavior to users and items respectively.

[0060]

[0061] S4: Personalized real interest and conformity behavior weight calculation module, introduces weight attention mechanism, uses multi-layer perception to obtain the attention scores of user real interest and conformity behavior embedding vectors and obtained by the previous module. The specific steps are as follows:

[0062] 1) Calculate the attention scores of and , for a given reason c∈{int,con}, the attention score calculation formula is as follows:

[0063]

[0064] 2) Normalization, use softmax function to normalize the attention scores of real interest and conformity behavior, the calculation formula is as follows:

[0065]

[0066] S5: Personalized recommendation score prediction module, based on the personalized weights of real interest and conformity behavior obtained by the previous module, weight the sum of interest score and conformity score to get the final personalized recommendation score The specific calculation formula is as follows:

[0067]

[0068] S6: Popularity-driven model optimization module, specifically comprising the following steps:

[0069] 1) Constructing loss function

[0070] BPR loss function is used for model training. Let the training set consist of triplets (u, i, j), where i and j are the items that have been interacted and not interacted by user u, respectively. According to the popularity of positive and negative samples i and j, the training set can be divided into two cases:

[0071] Case1: If the popularity of the positive sample i is greater than that of the negative sample j, then the corresponding training data set is denoted as D1. It can be considered that the comprehensive score of the positive sample is higher than that of the negative sample The conformity score of the positive sample is higher than that of the negative sample That is, the following inequality:

[0072]

[0073] Case2: If the popularity of the positive sample i is less than that of the negative sample j, then the corresponding training data set is denoted as D2. Since the comprehensive score of the positive sample is higher than that of the negative sample The conformity score of the positive sample is lower than that of the negative sample It can be deduced that the interest score of the positive sample is higher than that of the negative sample That is, the following inequality:

[0074]

[0075] Based on the above inequality, the following comprehensive loss function is used to optimize model training.

[0076]

[0077] 2) Item popularity-driven negative sampling

[0078] The popularity ISP based on item similarity calculated by the above module i is used as the basis for negative sampling to construct training data sets D1 and D2.

[0079] S7: Recommendation result acquisition module, used to acquire the recommendation list of target user u. The comprehensive prediction values of the items that have not been rated by the user are sorted, and the top k items are selected to form the personalized recommendation list of target user u.

[0080] Figure 3The proposed method was measured on three datasets across three metrics, and compared with other methods. The four metrics are: HR, NDCG, and PRU. Their calculation methods and measurement details are shown below:

[0081] HR is used to measure the accuracy of the recommendation list, measuring whether the actual recommended items in the test set appear on the top-k predicted recommendation list. The calculation method is as follows:

[0082]

[0083] Where hit(u) represents the proportion of the top k predicted recommended items for user u in the actual set of interacting items.

[0084] NDCG indicates the location of clicks by assigning higher scores to top-ranked interactions. It reflects the quality of the model's item ranking recommendations, and the calculation formula is as follows:

[0085]

[0086]

[0087] Among them, DCG u @K indicates recommending K items to user u to accumulate the benefits of their losses. k ∈{0,1} represents the relevance of the recommendation result for location idx. If the item is in the user's actual purchased item set, then rel idx The value is 1 otherwise. IDCG u @K is pressing rel. k The ideal DCG calculated on the descending list of recommendations u @K. K represents the length of the recommendation list.

[0088] PRU calculates the mean Spearman coefficient of item popularity and item ranking in each user's recommendation list, representing the correlation between item popularity and item ranking in the recommendation list, and is used to measure the popularity deviation of the recommendation list.

[0089]

[0090] Where rank(pop(I)) represents the ranking of the popularity of items in the recommendation list. u (I) represents the ranking of the item in the recommendation list, and SRC(·,·) represents the calculation of the Spearman correlation coefficient between the two variables.

Claims

1. A popularity-based recommendation model based on item similarity to decouple users' true interests and herd behavior, alleviate popularity bias, and improve recommendation quality, including the following: The data acquisition and preparation module is used to acquire and clean the data collected in the database to obtain the required raw scoring matrix; The project similarity and popularity calculation module is used to calculate project similarity and popularity. Starting from the fine-grained attributes of the project, it uses Kullback-Leibler (KL) divergence to calculate project similarity, dynamically defines "same type" projects through project similarity, and calculates the relative popularity of the project in its category, which is defined as ISP. It can obtain a ranking that represents the true popularity of the projects. The causal embedding and decoupling module, from a causal perspective, decouples user-item interaction behavior into genuine interest and conformity behavior, models it as a combination of genuine interest score and conformity behavior score, and assigns embedding vectors representing interest and conformity behavior to users and items respectively. The personalized genuine interest and herd behavior weight calculation module is designed with an adaptive weight attention mechanism to capture the weight of users' personalized genuine interests and herd behavior. The personalized recommendation score prediction module, based on the weights obtained from the previous module, combines the user's real interests and herd behavior scores to form the final personalized prediction score of the user for the project. Popularity-driven model optimization module for training model parameters, including loss function construction and negative sampling methods; This module provides a supervisory signal for learning the embedding vectors of genuine interest and herd mentality by introducing popularity information; The recommendation results acquisition module sorts the comprehensive predicted values ​​of unrated items by users and selects the top k items to generate a list of recommended items for users. The project similarity and popularity calculation module is used to calculate project popularity and project popularity, and the specific steps are as follows: S1: Calculate the similarity between items sim(i,j) using KL divergence; S2: Based on the calculated similarity, select the N items with the highest similarity to the target item i as its class item set S. i Next, calculate the target item i in S. i The relative popularity is calculated using the following formula: Among them, pop i The global popularity of project i, as defined in the traditional sense, is defined as the frequency of interaction among all users. The popularity-driven model optimization module specifically includes the following parts: S1: Loss function construction, using the BPR loss function for model training; The training set consists of triples (u, i, j), where i and j represent the items that user u has interacted with and those that haven't, respectively; Based on the popularity of positive and negative samples i and j, the training set can be divided into two cases: Case 1: If the popularity of positive sample i is greater than the popularity of negative sample j, then the corresponding training dataset is denoted as D1; ​​it can be considered that the overall score of positive samples... Higher than negative samples Conformity score of positive samples Higher than negative samples That is, the following inequality: Case 2: If the popularity of positive sample i is less than the popularity of negative sample j, then the corresponding training dataset is denoted as D2; since the overall score of positive samples... Higher than negative samples Conformity score of positive samples Lower than negative samples It can be deduced that the interest scores of positive samples Higher than negative samples That is, the following inequality: Based on the above inequalities, the following comprehensive loss function is designed to optimize model training: S2: Popularity based on project similarity ISP (Internet Service Provider Index) i As a basis for negative sampling, sampling is performed to construct training datasets D1 and D2, so that each training dataset can provide effective supervision signals to support the model in learning causal embeddings.

2. The popularity-based recommendation model based on item similarity as described in claim 1, characterized in that, The causal embedding and decoupling module, from a causal perspective, decouples user-item interaction behavior into genuine interest and conformity behavior, modeling it as a combination of genuine interest score and conformity behavior score; and assigns embedding vectors representing interest and conformity behavior to users and items respectively, helping the recommendation model learn more interpretable user and item features, as shown in the following formula: Where <·,·> denote the dot product; Let r be the embedding vectors of genuine interest and conformity, respectively; ui Rate item i for user u; Score for a specific reason.

3. The popularity-based recommendation model based on item similarity as described in claim 1, characterized in that, The personalized genuine interest and conformity behavior weight calculation module introduces a multilayer perceptron (MLP) to enhance the model's decoupling capability. This enables the model to capture complex feature relationships and patterns, thereby adaptively learning the optimal weight distribution to accurately reflect the user's genuine interests and conformity behavior. The specific implementation is as follows: S1: Given a reason c belonging to a set consisting of users' genuine interests and conformity behaviors, denoted as c∈{int,con}; for a specific reason c, its corresponding embedding The attention score of this reason is used as input to the MLP. Represented as: in, All of these are model parameters, where s represents the number of neurons in the hidden layer. S2: The attention score is normalized using the softmax function, calculated as follows:

4. The popularity-based recommendation model based on item similarity as described in claim 1, characterized in that, The personalized recommendation score prediction module predicts users' personalized ratings for items by combining the embedding vectors of user interests and conformity behavior. First, from a causal embedding perspective, the interaction between users and items is modeled as a weighted combination of interest scores and conformity behavior scores. Then, the learned embedding vectors are used to calculate the predicted ratings for each item, using the following formula:

5. The popularity-based recommendation model based on item similarity as described in claim 1, characterized in that, The recommendation result acquisition module is used to acquire the recommendation list for the target user u; The overall predicted values ​​of the unrated items by the user are sorted, and the top k items are selected to form a personalized recommendation list for the target user u.

Images