Universal recommendation framework for solving data sparsity

Abstract

The invention provides a universal recommendation framework for solving data sparsity, which comprises the following steps of: acquiring and cleaning data information from a database to obtain required information; generating a user interaction matrix T by using the original scoring matrix R of the user and the article, and then empowering the common interaction matrix by using popularity; generating a similarity matrix S by using the original scoring matrix R of the user and the article and adopting a similarity calculation method; the common interaction value Ti, j is compared with a reliability threshold value t set by the system, the similarity of the corresponding positions with the common interaction value lower than the threshold value is filtered, and the similarity value Si, j of the high-reliability position is reserved; filling the denoised similarity matrix by using a matrix decomposition model MF, and performing symmetry processing to generate a denoised and filled similarity matrix; combining the similarity value after denoising filling with the original score value to obtain a predicted value of the user to the article; and sorting the comprehensive predicted values of the non-scored items by the user, and selecting the first k items to generate a recommended item list of the user.

Description

Technical Field

[0001] The present invention relates to the field of recommendation technology, and in particular to a universal recommendation framework for solving data sparsity. Background Art

[0002] In recent years, with the popularization and in-depth application of the Internet and mobile Internet, the amount of information has increased dramatically. How to solve information overload and meet the personalized needs of users has become a current research hotspot. As the core of the recommendation system, recommendation algorithms have been widely used and studied, and collaborative filtering technology is one of the most successful recommendation algorithms currently used. In the recommendation algorithm, the most important thing is to calculate the similarity between users or items, which directly determines the quality of the recommendation system.

[0003] Many similarity measurement methods have been proposed to improve recommendation performance. However, most memory-based methods directly use the calculated similarity for calculation without considering the reliability and effectiveness of the similarity results, resulting in poor recommendation performance when the data sparsity becomes low. Therefore, it is particularly important to design a general recommendation framework that solves data sparsity to ensure the accuracy of recommendations under various sparse data. Summary of the invention

[0004] In view of the above problems, the present invention proposes a universal recommendation framework for solving data sparsity, which is used to solve the recommendation under sparse data and improve the recommendation quality, including the following:

[0005] Data collection and preparation module, used to acquire and clean the data collected in the database to obtain the required information;

[0006] The interaction matrix generation module is used to generate the reliability threshold. It uses the original rating matrix R of users and items to generate the user interaction matrix T, and then uses popularity to weight the common interaction matrix. After weighting, each value in the matrix is ​​the common interaction value to be screened.

[0007] A similarity matrix generation module is used to generate similarity values, using the original rating matrix R of users and items to generate a similarity matrix S using a similarity calculation method;

[0008] The reliability threshold filtering module is used to filter unreliable similarity values. The interaction value output by the interaction matrix generation module is compared with the reliability threshold set by the system, and the similarity of the corresponding positions whose common interaction value is lower than the threshold is filtered to retain the similarity value with higher reliability.

[0009] The matrix decomposition filling module is used to fill the sparse similarity matrix after denoising. The matrix decomposition model MF is used to fit the denoised similarity matrix, predict the similarity values ​​between different users, and finally perform symmetry processing to generate the denoised and filled similarity matrix;

[0010] A comprehensive prediction value calculation module is used to combine the denoised and filled similarity values ​​and the original rating values ​​to obtain the user's prediction value for the item;

[0011] The recommendation result acquisition module is used to sort the comprehensive prediction values ​​of users for unrated items and select the top k items to generate a recommended item list for the user.

[0012] The interaction matrix generation module of the present invention is used to generate a reliability threshold, use the original rating matrix R of users and items to generate a user interaction matrix T, and then use popularity to weight the common interaction matrix. After weighting, each value in the matrix is ​​a common interaction value to be screened. The weighting rules are as follows:

[0013]

[0014]

[0015] Among them: pop k represents the number of users who have interacted with the kth item, and sf represents the scaling factor.

[0016] The similarity matrix generation module of the present invention is used to generate similarity values. The similarity matrix S can be generated by using the original rating matrix R of users and items and adopting a similarity calculation method. If Pearson similarity is used as the similarity calculation method, the specific calculation formula is as follows:

[0017]

[0018] Where: I u is the set of items rated by user u, is the average rating of all rated items by user u, r ui represents the rating of user u on item i.

[0019] The reliability threshold filtering module described in the present invention is used to filter unreliable similarity values. The common interaction value output by the interaction matrix generation module is compared with the reliability threshold t set by the system, and the similarities of corresponding positions with common interaction values ​​lower than the threshold are filtered, and the similarity values ​​of positions with higher reliability are retained.

[0020] The matrix decomposition filling module of the present invention is used to fill the sparse similarity matrix after denoising. The matrix decomposition model MF is used to fit the denoised similarity matrix, predict the similarity values ​​between different users, and finally perform symmetry processing to generate the denoised and filled similarity matrix. Specifically, it includes the following steps:

[0021] S1: Use matrix decomposition to fit the similarity value as follows:

[0022]

[0023] In order to train the latent factor matrices P and Q, it is necessary to minimize the regularized square error to obtain the optimal value, that is:

[0024]

[0025] Where: N is the set of users; S i,j is the value of similarity; ||·|| 2 is the Frobenius norm of the matrix; λ is a regulatory variable used to control the regularization strength to prevent overfitting of the score;

[0026] S2: Symmetric processing, specifically using the following formula for symmetric processing:

[0027]

[0028] The comprehensive prediction value calculation module of the present invention is used to combine the similarity value after denoising and filling with the original score value to obtain the user's prediction value for the item. The specific prediction formula is as follows:

[0029]

[0030] Where: N u represents the set of nearest neighbors of user u, is the average rating of all rated items by user u, r ui represents the rating of item i by user u, and sim(u,v) represents the similarity value between user u and v.

[0031] The recommendation result acquisition module of the present invention 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.

[0032] The present invention proposes a universal recommendation framework for solving data sparsity. It only uses the interaction relationship and rating information between users and items without other auxiliary information to obtain a reliable and effective similarity matrix. Therefore, the proposed framework has good versatility and scalability. At the same time, the proposed framework can improve the recommendation accuracy of traditional memory-based recommendation methods and maintain the recommendation accuracy on data with different sparsity levels by generating a dense and reliable similarity matrix to predict ratings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the system process of the present invention;

[0034] Figure 2 It is a system framework diagram of the present invention;

[0035] Figure 3 This is the improvement of MAE value, HR value and NDCG value using the framework of the present invention under three data sets. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, a general recommended framework for solving data sparsity proposed by the present invention is described in detail below 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 used to limit the present invention. Changes, modifications, additions, modifications or substitutions made by ordinary technicians in the technical field within the scope of the present invention should be included in the scope of the claims of the present invention.

[0037] Figure 1 It is a schematic diagram of a process of a general recommended framework for solving data sparsity in the present invention. Figure 1 It can be seen that the general recommendation framework for solving data sparsity proposed in the present invention includes obtaining and cleaning data information from a database to obtain required information; using the original rating matrix R of users and items to generate a user interaction matrix T, and then using popularity to weight the common interaction matrix; using the original rating matrix R of users and items and using a similarity calculation method to generate a similarity matrix S; by comparing the common interaction value with the reliability threshold t set by the system, the similarity of the corresponding position with the common interaction value lower than the threshold is filtered, and the similarity value S of the position with higher reliability is retained i,j ; Use the matrix decomposition model MF to fill the denoised similarity matrix and perform symmetry processing to generate the denoised and filled similarity moment; combine the denoised and filled similarity values ​​with the original rating values ​​to obtain the user's prediction value for the item; sort the user's comprehensive prediction values ​​for the unrated items, and select the top k items to generate the user's recommended item list.

[0038] Figure 2It is a framework diagram of a general recommended framework for solving data sparsity in the present invention. Figure 2 It can be seen that the general recommendation framework for solving data sparsity proposed by the present invention includes a data collection and preparation module, an interaction matrix generation module, a similarity matrix generation module, a reliability threshold filtering module, a matrix decomposition filling module, a comprehensive prediction value calculation module, and a recommendation result acquisition module. The data collection and preparation module is used to extract data from the database and perform operations such as cleaning and preprocessing on the data to obtain the required information; the interaction matrix generation module uses the original rating matrix R of users and items to generate the user interaction matrix T, and then uses popularity to weight the common interaction matrix; the similarity matrix generation module uses a similarity calculation method on the original rating matrix R of users and items to generate a similarity matrix S; the reliability threshold filtering module uses the common interaction value T i,j Compare it with the reliability threshold t set by the system, filter the similarities of the corresponding positions with common interaction values ​​lower than the threshold, and retain the similarity values ​​S of the positions with higher reliability i,j ; The matrix decomposition and filling module uses the matrix decomposition model MF to fit the denoised similarity matrix, predicts the similarity values ​​between different users, and finally performs symmetry processing to generate the denoised and filled similarity matrix; the comprehensive prediction value calculation module combines the denoised and filled similarity values ​​and the original rating values ​​to obtain the user's prediction value for the item; the recommendation result acquisition module sorts the user's comprehensive prediction values ​​for the unrated items, and selects the top k items to generate the user's recommended item list.

[0039] Further, the following situation is used as an example to illustrate:

[0040] There are M users 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, U(i,j) represents the set of users who have rated both item i and item j, i.e., the common rating item; U i represents the set of all users who have evaluated item i.

[0041] First, the system data is acquired and cleaned through the data acquisition and preparation module to obtain the variables in the hypothesis. The specific implementation steps of the proposed deep neural network recommendation system with predictive reliability are as follows:

[0042] S1: The data collection and preparation module obtains the required data information from the database, including user ID, project ID and rating information;

[0043] S2: The interaction matrix generation module uses the original rating matrix R of users and items to generate the user interaction matrix T, and then uses popularity to weight the common interaction matrix;

[0044] Taking the MovieLens dataset as an example, there are M users, and the generated interaction count matrix is ​​T M×M , where T ij represents the number of common interactions between user i and user j, and then uses the following formula T ij To empower:

[0045]

[0046] Among them pop k represents the number of users who have interacted with the kth item, and sf represents the scaling factor.

[0047] S3: Similarity matrix generation module is based on the original rating matrix R of users and items. M×N A similarity calculation method is used to generate the similarity matrix S M×M , such as using Pearson similarity to calculate the similarity between users;

[0048] S4: Reliability threshold filtering module through common interaction value T i,j Compare it with the reliability threshold t set by the system, filter the similarities of the corresponding positions with common interaction values ​​lower than the threshold, and retain the similarity values ​​S of the positions with higher reliability i,j ;

[0049] S5: The matrix decomposition filling module uses the matrix decomposition model MF to fit the denoised similarity matrix, predict the similarity values ​​between different users, and finally perform symmetry processing to generate the denoised and filled similarity matrix. Use matrix decomposition to fit the similarity value as follows:

[0050]

[0051] In order to train the latent factor matrices P and Q, it is necessary to minimize the regularized square error to obtain the optimal value, that is:

[0052]

[0053] Where: N is the set of users; S i,j is the value of similarity; ||·|| 2 is the Frobenius norm of the matrix; λ is a regulatory variable used to control the regularization strength to prevent overfitting of the score;

[0054] Symmetric processing, specifically using the following formula for symmetric processing:

[0055]

[0056] S6: The comprehensive prediction value calculation module combines the denoised and filled similarity values ​​with the original rating values ​​to obtain the user's prediction value for the item.

[0057] S7: Recommendation result acquisition module, used to obtain the recommendation list of target user u. Sort the comprehensive prediction values ​​of the user's unrated items, and select the top k items to form a personalized recommendation list for target user u.

[0058] Figure 3 The proposed method is measured in two aspects on three datasets and compared with other methods. The three metrics are: mean absolute error (MAE), hit rate (HR) and normalized discounted cumulative gain (NDCG). Their calculation methods and measurement contents are as follows:

[0059] Mean absolute error (MAE): Prediction accuracy is an important reference standard for measuring the performance of collaborative filtering algorithms. The mean absolute error is often used for measurement. The calculation formula is:

[0060]

[0061] Where: m represents the number of users in the recommendation system, n is the number of items predicted by user u; r ui and p ui are the actual rating and predicted rating of item i by user u, respectively.

[0062] Hit rate (HR): used to measure whether the actual recommended items in the test set are presented in the top-k predicted recommendation list, which indicates the item recommendation ability of the model. The calculation method is as follows:

[0063]

[0064] Where: m represents the number of users in the recommendation system, hits(i) represents the proportion of the top k predicted recommended items of the i-th user in the actual recommended item set.

[0065] Normalized Discounted Cumulative Gain (NDCG): This shows the quality of the model’s item ranking recommendations by assigning higher scores to top-ranked clicks to account for the click’s position. The calculation formula is as follows:

[0066]

[0067] Where DCG@k and IDCG@k represent the discounted cumulative gain and ideal DCG, respectively. p Used to show the recommendation relevance of the item at position p, that is, if the predicted recommended item appears in the actual recommendation list, rel p =1 if the predicted or actual rating of an item exceeds the median of the rating interval, otherwise it is 0. It is worth noting that we assume that the item is recommended if its predicted or actual rating exceeds the median of the rating interval.

Claims

1. A general recommendation framework for solving data sparsity, which is used to solve the recommendation under sparse data and improve the recommendation quality, including the following: Data collection and preparation module, used to acquire and clean the data collected in the database to obtain the required information; The interaction matrix generation module is used to generate the reliability threshold. It uses the original rating matrix R of users and items to generate the user interaction matrix T, and then uses popularity to weight the common interaction matrix. After weighting, each value in the matrix is ​​the common interaction value to be screened. A similarity matrix generation module is used to generate similarity values, using the original rating matrix R of users and items to generate a similarity matrix S using a similarity calculation method; The reliability threshold filtering module is used to filter unreliable similarity values. The interaction value output by the interaction matrix generation module is compared with the reliability threshold set by the system, and the similarity of the corresponding positions whose common interaction value is lower than the threshold is filtered to retain the similarity value with higher reliability. The matrix decomposition filling module is used to fill the sparse similarity matrix after denoising. The matrix decomposition model MF is used to fit the denoised similarity matrix, predict the similarity values ​​between different users, and finally perform symmetry processing to generate the denoised and filled similarity matrix; A comprehensive prediction value calculation module is used to combine the denoised and filled similarity values ​​and the original rating values ​​to obtain the user's prediction value for the item; The recommendation result acquisition module is used to sort the comprehensive prediction values ​​of the user for the unrated items and select the top k items to generate the user's recommended item list.

2. A general recommendation framework for solving data sparsity as claimed in claim 1, characterized in that: The interaction matrix generation module is used to generate a reliability threshold, use the original rating matrix R of users and items to generate a user interaction matrix T, and then use popularity to weight the common interaction matrix. After weighting, each value in the matrix is ​​a common interaction value to be screened. The weighting rules are as follows: Among them: pop k represents the number of users who have interacted with the kth item, and sf represents the scaling factor.

3. A general recommendation framework for solving data sparsity as claimed in claim 1, characterized in that: The similarity matrix generation module is used to generate similarity values, and uses the original rating matrix R of users and items and adopts a similarity calculation method to generate a similarity matrix S. If Pearson similarity is used as the similarity calculation method, the specific calculation formula is as follows: Where: I u is the set of items rated by user u, is the average rating of all rated items by user u, r ui represents the rating of user u on item i.

4. A general recommendation framework for solving data sparsity as claimed in claim 1, characterized in that: The reliability threshold filtering module is used to filter unreliable similarity values, and the common interaction value T output by the interaction matrix generation module is i,j Compare it with the reliability threshold t set by the system, filter the similarities of the corresponding positions with common interaction values ​​lower than the threshold, and retain the similarity values ​​S of the positions with higher reliability i,j .

5. A general recommendation framework for solving data sparsity as claimed in claim 1, characterized in that: The matrix decomposition filling module is used to fill the sparse similarity matrix after denoising. The matrix decomposition model MF is used to fit the denoised similarity matrix, predict the similarity values ​​between different users, and finally perform symmetry processing to generate the denoised and filled similarity matrix. Specifically, it includes the following steps: S1: Use matrix decomposition to fit the similarity value as follows: In order to train the latent factor matrices P and Q, it is necessary to minimize the regularized square error to obtain the optimal value, that is: Where: N is the set of users; S i,j is the value of similarity; ||·|| 2 is the Frobenius norm of the matrix; λ is a regulatory variable used to control the regularization strength to prevent overfitting of the score; S2: Symmetric processing, specifically using the following formula for symmetric processing:

6. A general recommendation framework for solving data sparsity as claimed in claim 1, characterized in that: The comprehensive prediction value calculation module is used to combine the similarity value after denoising and filling with the original score value to obtain the user's prediction value for the item. The specific prediction formula is as follows: Where: N u represents the set of nearest neighbors of user u, is the average rating of all rated items by user u, r ui represents the rating of item i by user u, and sim(u,v) represents the similarity value between user u and v.

7. A general recommendation framework for solving data sparsity 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