Project similarity-based popularity depolarization recommendation model

Abstract

The invention provides a popularity depolarization recommendation model based on item similarity, which comprises the following steps of: extracting user item score data from a database and preprocessing the user item score data to obtain an original score matrix; calculating the similarity among the items based on KL divergence, and calculating the popularity (ISP) based on item similarity based on the similarity; based on the causal theory, the interaction evaluation of the user and the project is decomposed and coupled into two parts of real interest and conformity behavior, and corresponding embedded vectors are distributed for the user and the project; by introducing a weight attention mechanism, the interests of the users and the weights of the crowd behaviors are learned in a self-adaptive mode; calculating a final personalized recommendation score based on the obtained weight weighted summation of the scores of the two parts; a training data set is divided through a negative sampling method based on ISP, and a loss function is constructed to supervise learning of a decoupling embedded vector; and performing score value prediction and sorting on the items which are not scored by the user, and selecting the first k items to form a personalized recommendation list.

Description

Technical Field

[0001] The present invention relates to the field of recommendation technology, and in particular to a popularity debiasing recommendation model based on item similarity. Background Art

[0002] Recommender systems play a vital role in addressing information overload, improving user experience, and increasing user engagement, making them an indispensable technology in the information age. Existing recommendation models typically model user interests based on observed user-item interaction data. This observed data can also be influenced by user behavior that follows popular items. Consequently, traditional recommendation models are often influenced by item popularity when learning user interests, making it difficult to accurately capture users' true interests. This leads to reduced recommendation accuracy and the potential for popularity bias.

[0003] To address the problem of popularity bias, some existing debiasing methods reduce bias through techniques such as data preprocessing, regularization, and reweighting training samples. Existing methods still face the following problems when dealing with the decoupling of interest and herd behavior: First, in observational data, user interests and herd behavior are often mixed together, so finding a reasonable signal to distinguish between the two is crucial for effective training on the original observational data. However, these methods are based on traditional global item popularity and calculate popularity only through interaction frequency, failing to consider the actual popularity of items at a more granular level. This may lead to inaccurate distinctions between interest and herd behavior, thus affecting the training effect of the model. Second, existing debiased recommendation models model user-item interaction behavior as a simple sum of interest scores and herd behavior scores, which makes it difficult to represent users' personalized behavior, affecting the model's expressiveness and, in turn, the recommendation performance.

[0004] Therefore, a popularity-debiased recommendation model based on item similarity is designed. By introducing a popularity calculation method based on item similarity to generate a more reasonable negative sampling signal, and introducing a weighted attention mechanism to learn the attention scores of users' real interests and herd behavior, it can adaptively capture the weights of specific users' real interests and herd behavior, which is of great significance for improving the recommendation model's ability to remove popularity bias. Summary of the Invention

[0005] To address the above issues, this paper proposes a popularity-based debiased recommendation model based on item similarity. By generating a more reasonable negative sampling signal based on item similarity's popularity, this model effectively distinguishes user interests from herd behavior during training. Furthermore, this paper designs an adaptive weighted attention mechanism to capture user personalized behavior regarding interests and herd behavior. Specifically, it includes the following:

[0006] The data collection and preparation module is used to obtain and clean the data collected in the database to obtain the required original scoring matrix.

[0007] The project similarity and popularity calculation module is used to calculate project similarity and project popularity. Starting from the fine-grained attributes of the project, it uses Kullback-Leibler (KL) divergence to calculate project similarity, dynamically defines "similar" projects through project similarity, and calculates the relative popularity of the project in its category, defined as ISP, which can obtain a ranking that represents the true popularity of the project.

[0008] The causal embedding and decoupling module uses causal embedding technology to learn users' interests and herd behavior, and decouples user-item interaction behavior into real interests and herd behavior.

[0009] The personalized real interest and herd behavior weight calculation module designs an adaptive weighted attention mechanism to capture the user's personalized real interest and herd behavior weight.

[0010] The personalized recommendation score prediction module combines the real interest score and the herd mentality score to generate the user's final prediction value for the item.

[0011] The popularity-driven model optimization module is used to train model parameters, including the construction of loss functions and negative sampling methods. By introducing popularity information, this module provides supervision signals for the learning of true interest and herd mentality embedding vectors.

[0012] The recommendation result acquisition module ranks the comprehensive prediction values ​​of users for unrated items and selects the top k items to generate a list of recommended items for the user.

[0013] The project similarity and popularity calculation module of the present invention considers fine-grained project attributes and is used to calculate project similarity and popularity. The specific steps are as follows:

[0014] 1) Calculate the similarity sim(i,j) between items using KL divergence. The calculation formula is as follows:

[0015]

[0016] Among them, ρ ir , ρ jr is the frequency of occurrence of score r in item i and item j.

[0017] 2) Select the N projects with the highest similarity to the target project i as its similar project set S i .

[0018] 3) Calculate the target item i in S i The relative popularity in is calculated as follows:

[0019]

[0020] Among them, pop i is the global item popularity of item i under the traditional definition, which is defined as the interaction frequency among all users.

[0021] The causal embedding and decoupling module described in this paper decouples user-item interaction behavior from a causal perspective into true interest and conformity, modeling it as a combination of true interest scores and conformity scores. It also assigns embedding vectors representing interest and conformity to users and items, respectively, helping the recommendation model learn more interpretable user and item features. This is shown in the following formula:

[0022]

[0023] Where <·,·> represents the dot product; are respectively represented as embedding vectors of real interest and herd mentality; r ui is the rating of item i by user u; Score for a specific reason.

[0024] The personalized real interest and herd behavior weight calculation module described in this invention introduces a multi-layer perceptron (MLP) to enhance the model's decoupling capabilities. This enables the model to capture complex feature relationships and patterns, thereby adaptively learning the optimal weight distribution and accurately reflecting the user's real interests and herd behavior. The specific implementation is as follows:

[0025] 1) Given a reason c, it belongs to a set consisting of the user’s real interests and herd behavior, denoted as c∈{int,con}. For a specific reason c, its corresponding embedding As input to the MLP, the attention score of the reason Expressed as:

[0026]

[0027] in, are all model parameters, and s represents the number of neurons in the hidden layer.

[0028] 2) Use the softmax function to normalize the attention score. The calculation formula is as follows:

[0029]

[0030] The personalized recommendation score prediction module described in this invention predicts a user's personalized rating for an item by combining the embedding vectors of user interests and conformity. First, from the perspective of causal embedding, the interaction between the user and the item is modeled as a weighted combination of interest scores and conformity scores. The learned embedding vectors are then used to calculate the user's predicted rating for each item. The calculation formula is as follows:

[0031]

[0032] The popularity-driven model optimization module of the present invention specifically includes the following parts:

[0033] 1) Loss function construction

[0034] The BPR loss function is used for model training. Assume that the training set consists of triples (u, i, j), where i and j are the items that user u has interacted with and has not interacted with, respectively. Based on the popularity of positive and negative samples i and j, the training set can be divided into two cases:

[0035] Case 1: If the popularity of positive sample i is greater than the popularity of negative sample j, the corresponding training data set is recorded as D1. It can be considered that the comprehensive score of the positive sample Higher than negative samples The conformity score of the positive sample Higher than negative samples That is the following inequality:

[0036]

[0037] Case 2: If the popularity of positive sample i is less than the popularity of negative sample j, the corresponding training data set is recorded as D2. Higher than negative samples The conformity score of the positive sample Lower than negative samples It can be deduced that the interest score of the positive sample is Higher than negative samples That is the following inequality:

[0038]

[0039] Based on the above inequality, the following comprehensive loss function is designed to optimize model training.

[0040]

[0041] 2) Negative sampling driven by item popularity

[0042] The popularity ISP based on item similarity calculated by the above module iAs the basis for negative sampling, sampling is performed to construct training data sets D1 and D2, so that each training data set can provide an effective supervision signal to support the model to learn causal embedding.

[0043] The recommendation result acquisition module of the present invention ranks the comprehensive scores of the users for the unrated items and selects the top k items to generate a recommended item list for the user.

[0044] This paper proposes a popularity-debiased recommendation model based on item similarity. From a causal perspective, it decouples user-item interactions into true interest and herd behavior. It also introduces an adaptive weighted attention mechanism to dynamically adjust the weights of true interest and herd behavior based on the user's personalized behavior. To ensure model optimization, the paper constructs a reasonable loss function and calculates item popularity based on KL divergence. This provides a more reasonable supervisory signal for model training, ensuring that user interests and herd behavior are embedded in the learning process, thereby improving the accuracy and debiasing capabilities of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the model process of the present invention;

[0046] Figure 2 It is a model framework diagram of the present invention;

[0047] Figure 3 This is the comparison of HR@5, NDCG@5, HR@10, NDCG@10, and PRU@10 of the present invention with other methods under three data sets. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clear, the following is a detailed description of a neural collaborative filtering recommendation model that considers multi-dimensional user preferences proposed by the present invention in conjunction with the accompanying drawings. It should be understood that the specific implementation method described herein is only used to explain the present invention and is not intended to limit the present invention. Changes, modifications, additions, modifications or substitutions made by ordinary technicians in this technical field within the scope of the present invention should be included in the scope of the claims of the present invention.

[0049] Figure 1This is a process diagram of a popularity debiased recommendation model based on project similarity of the present invention. First, the user-project rating data is extracted from the database and preprocessed to obtain the original rating matrix. Then, the similarity between projects is calculated based on KL divergence, and the popularity based on project similarity (ISP) is calculated based on this similarity. At the same time, based on causal theory, the interaction rating between users and projects is decoupled into two parts: real interest and herd behavior, and corresponding embedding vectors are assigned to users and projects. On this basis, by introducing a weighted attention mechanism, the weights of users' interests and herd behavior are adaptively learned. Next, the final personalized recommendation score is calculated based on the weighted sum of the two scores obtained. The training data set is divided by the ISP-based negative sampling method, and the constructed loss function is used for model optimization. After the model training is completed, the rating values ​​of the projects that have not been rated by the user are predicted and sorted, and the top k projects are selected to form a personalized recommendation list.

[0050] Figure 2 This is a framework diagram of the popularity-debiased recommendation model based on item similarity. It includes a data collection and preparation module, an item similarity and popularity calculation module, a causal embedding and decoupling module, a personalized real interest and herd behavior weight calculation module, a personalized recommendation score prediction module, a popularity-driven model optimization module, and a recommendation result acquisition module. The data collection and preparation module extracts data from the database and performs operations such as data cleaning and preprocessing to obtain the required raw user-item rating matrix. The item similarity and popularity calculation module calculates the KL similarity and popularity between items, ensuring that the popularity accurately reflects the popularity of the items. The causal embedding and decoupling module, based on causal theory, decouples the interaction ratings between users and items into two components: "real interest" and "herd behavior," and assigns corresponding embedding vectors to each user and item. The personalized real interest and herd behavior weight calculation module introduces a weighted attention mechanism and uses a multi-layer perceptron to calculate the attention weight of each user's interest preference and herd behavior. The personalized recommendation score prediction module generates a final personalized recommendation score based on the weighted sum of the personalized interest and herd behavior weights obtained in the previous step, providing a basis for subsequent recommendation list sorting. The recommendation result acquisition module obtains the recommendation list for the target user u. The comprehensive predicted values ​​of the user's unrated items are sorted, and the top k items are selected to form the personalized recommendation list for the target user u.

[0051] Further, the following situation is used as an example to illustrate:

[0052] There are M users U in the MovieLens dataset U={u1,u2,...,u M} and N items I={i1,i2,...,i N}, the rating domain is {1,2,3,4,5}, and the user-item rating matrix is ​​[R] M×N Indicates; r ui Represents the user's rating value on item i.

[0053] First, the system data is acquired and cleaned through the data collection and preparation module to obtain the variables in the hypothesis. The specific implementation steps of the proposed popularity debiasing recommendation model based on item similarity are as follows:

[0054] S1: The data collection and preparation module obtains the required data information from the database, including user ID, project ID and rating information.

[0055] S2: The project similarity and popularity calculation module is used to calculate project similarity and project popularity. The specific steps are as follows:

[0056] 1) Calculate the similarity sim(i,j) between items using KL divergence. The calculation formula is as follows:

[0057]

[0058] Among them, ρ ir , ρ jr is the frequency of occurrence of score r in item i and item j.

[0059] 2) Define the set S of similar projects of project i i Based on the calculated similarity, select the N projects with the highest similarity to the target project i as its similar project set S i .

[0060] 3) Calculate the target item i in S i The relative popularity in is calculated as follows:

[0061]

[0062] Among them, pop i is the global item popularity of item i under the traditional definition, which is defined as the interaction frequency among all users.

[0063] S3: Causal embedding and decoupling module decouples user-item ratings into real interest scores based on causal theory and conformity scores , and assign embedding vectors representing true interests and herd behavior to users and items respectively.

[0064]

[0065] S4: Personalized real interest and herd behavior weight calculation module, introduces the weighted attention mechanism, and uses a multi-layer perceptron to obtain the embedding vector of the user's real interest and herd behavior obtained in the previous module and The specific steps are as follows:

[0066] 1) Calculation and For a given reason c∈{int,con}, the attention score is calculated as follows:

[0067]

[0068] 2) Normalization: Use the softmax function to normalize the attention scores of true interest and herd behavior. The calculation formula is as follows:

[0069]

[0070] S5: Personalized recommendation score prediction module, based on the real interest and herd behavior personalized weights obtained in the previous module, performs weighted summation of the interest score and the herd score to obtain the final personalized recommendation score The specific calculation formula is as follows:

[0071]

[0072] S6: Popularity-driven model optimization module, specifically including the following steps:

[0073] 1) Constructing the loss function

[0074] The BPR loss function is used for model training. Assume that the training set consists of triples (u, i, j), where i and j are the items that user u has interacted with and has not interacted with, respectively. Based on the popularity of positive and negative samples i and j, the training set can be divided into two cases:

[0075] Case 1: If the popularity of positive sample i is greater than the popularity of negative sample j, the corresponding training data set is recorded as D1. It can be considered that the comprehensive score of the positive sample Higher than negative samples The conformity score of the positive sample Higher than negative samples That is the following inequality:

[0076]

[0077] Case 2: If the popularity of positive sample i is less than the popularity of negative sample j, the corresponding training data set is recorded as D2. Higher than negative samples The conformity score of the positive sample Lower than negative samples It can be deduced that the interest score of the positive sample is Higher than negative samples That is the following inequality:

[0078]

[0079] Based on the above inequality, the following comprehensive loss function is used to optimize model training.

[0080]

[0081] 2) Negative sampling driven by item popularity

[0082] The popularity ISP based on item similarity calculated by the above module i As the basis for negative sampling, sampling is performed to construct training data sets D1 and D2.

[0083] S7: Recommendation result acquisition module, used to obtain the recommendation list of target user u. The comprehensive prediction values ​​of the user's unrated items are sorted, and the top k items are selected to form a personalized recommendation list for target user u.

[0084] Figure 3 The proposed method was measured on three datasets and compared with other methods in three aspects. The four metrics are HR, NDCG, and PRU. Their calculation methods and measurement contents are as follows:

[0085] 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:

[0086]

[0087] Among them, hit(u) represents the proportion of the top k predicted recommended items of user u in the actual interaction item set.

[0088] NDCG accounts for the location of clicks by assigning higher scores to top-ranked interactions. It shows the quality of the model's item ranking recommendations and is calculated as follows:

[0089]

[0090] Among them, DCG u @K represents the cumulative loss of recommending K items to user u, rel k ∈{0,1} represents the relevance of the recommendation result at position idx. If the item is in the user's actual purchase item set, then rel idxThe value is 1, otherwise it is 0. u @K is pressing rel k Ideal DCG calculated on the descending recommendation list u @K. K represents the length of the recommendation list.

[0091] PRU calculates the mean of the Spearman coefficient of item popularity and item ranking in each user's recommendation list, which represents the correlation between item popularity and the ranking of recommendation list items and is used to measure the popularity deviation of the recommendation list.

[0092]

[0093] Among them, rank(pop(I)) represents the ranking of the popularity of the items in the recommendation list, rank u (I) represents the ranking of the item in the recommendation list, and SRC(·,·) represents the calculation of the Spearman correlation coefficient between two variables.

Claims

1. A popularity debiasing recommendation model based on item similarity is used to decouple users' real interests and herd behavior, alleviate popularity bias, and improve recommendation quality, including the following: The data collection and preparation module is used to obtain and clean the data collected in the database to obtain the required original scoring matrix; The project similarity and popularity calculation module is used to calculate project similarity and project popularity. Starting from the fine-grained attributes of the project, the Kullback-Leibler (KL) divergence is used to calculate project similarity, and the "same type" projects are dynamically defined through project similarity. The relative popularity of the project in its category is calculated, which is defined as ISP. The ranking that represents the real popularity of the project can be obtained. The causal embedding and decoupling module decouples user-item interaction behaviors into real interests and conformity behaviors from a causal perspective, models them as a combination of real interest scores and conformity scores, and assigns embedding vectors representing interests and conformity behaviors to users and items, respectively. The personalized real interest and herd behavior weight calculation module designs an adaptive weighted attention mechanism to capture the user's personalized real interest and herd behavior weight; The personalized recommendation score prediction module, based on the weights obtained in the previous module, combines the user's real interests and herd behavior scores to form the final user's personalized prediction score for the project; Popularity-driven model optimization module, which is used to train model parameters, including loss function construction and negative sampling methods. This module provides supervision signals for learning real interest and herd mentality embedding vectors by introducing popularity information; The recommendation result acquisition module sorts the comprehensive prediction values ​​of users for unrated items and selects the top k items to generate a recommended item list for the user.

2. The popularity debiasing recommendation model based on item similarity as claimed in claim 1, characterized in that: The project similarity and popularity calculation module considers fine-grained project attributes and is used to calculate project similarity and project popularity. The specific steps are as follows: S1: Calculate the similarity sim(i,j) between items through KL divergence. The calculation formula is as follows: Among them, ρ ir , jr is the frequency of occurrence of score r in item i and item j. S2: Based on the calculated similarity, select the N projects with the highest similarity to the target project i as its similar project set S i . Again, calculate the target item i in S i The relative popularity in is calculated as follows: Among them, pop i is the global item popularity of item i under the traditional definition, defined as the interaction frequency among all users.

3. The popularity debiasing recommendation model based on item similarity as claimed in claim 1, characterized in that: The causal embedding and decoupling module decouples user-item interaction behaviors into real interests and conformity behaviors from a causal perspective, and models them as a combination of real interest scores and conformity scores. It also assigns embedding vectors representing interests and conformity behaviors to users and items, respectively, to help the recommendation model learn more interpretable user and item features. This is shown in the following formula: Where <·,·> represents the dot product; They are respectively represented as the embedding vectors of real interest and herd mentality; r ui is the rating of item i by user u; Score for a specific reason.

4. The popularity debiasing recommendation model based on item similarity as claimed in claim 1, characterized in that: The personalized real interest and herd behavior weight calculation module introduces a multi-layer perceptron (MLP) to enhance the model decoupling capability. This enables the model to capture complex feature relationships and patterns, thereby adaptively learning the optimal weight distribution and accurately reflecting the user's real interest and herd behavior. The specific implementation is as follows: S1: Given a reason c, it belongs to a set consisting of the user's real interests and herd behavior, denoted as c∈{int,con}. For a specific reason c, its corresponding embedding As input to the MLP, the attention score of the reason It is expressed as: in, are all model parameters, and s represents the number of neurons in the hidden layer. S2: Use the softmax function to normalize the attention score. The calculation formula is as follows:

5. The popularity debiasing recommendation model based on item similarity as claimed in claim 1, characterized in that: The personalized recommendation score prediction module predicts the personalized score of the user for the project by combining the embedding vectors of the user's interest and herd behavior. First, the interaction between the user and the project is modeled as a weighted combination of the interest score and the herd behavior score from the perspective of causal embedding, and then the user's predicted score for each project is calculated using the learned embedding vector. The calculation formula is as follows:

6. The popularity debiasing recommendation model based on item similarity according to claim 1, wherein the popularity-driven model optimization module specifically comprises the following parts: S1: Loss function construction. The BPR loss function is used for model training. Assume that the training set consists of triples (u, i, j), where i and j are the items that user u has interacted with and has not interacted with, respectively. According to 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, the corresponding training data set is recorded as D1. It can be considered that the comprehensive score of the positive sample Higher than negative samples The 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, the corresponding training data set is recorded as D2. Higher than negative samples The conformity score of positive samples Lower than negative samples It can be deduced that the interest score of the positive sample is Higher than negative samples That is the following inequality: Based on the above inequality, the following comprehensive loss function is designed to optimize model training. S2: Negative sampling driven by project popularity. The popularity ISP based on project similarity calculated by the above module i As the basis for negative sampling, sampling is performed to construct training data sets D1 and D2, so that each training data set can provide an effective supervision signal to support the model to learn causal embedding.

7. The popularity debiasing recommendation model based on item similarity as claimed in claim 1, characterized in that: The recommendation result acquisition module is used to obtain the recommendation list of the target user u. The comprehensive prediction values ​​of the user's unrated items are sorted, and the top k items are selected to form a personalized recommendation list of the target user u.

Images