An Improved Width Collaborative Filtering Method Based on Bilateral Feature Alignment
An improved width collaborative filtering method with bilateral feature alignment is adopted. The rating is decomposed from the user side and the item side. A gated fusion factor α is introduced for feature alignment and fusion, which solves the problem of inconsistent feature distribution caused by differences in user rating criteria and improves the accuracy and adaptability of the recommendation system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-30
AI Technical Summary
Existing collaborative filtering methods based on width models have failed to effectively address the inconsistency in feature distribution caused by differences in user rating criteria. They lack personalized information extraction and fusion strategies from both user and project perspectives, resulting in insufficient recommendation accuracy and universality, and making it difficult to adapt to data scenarios with different sparsity levels.
An improved width collaborative filtering method based on bilateral feature alignment is adopted. By decomposing the score from both the user and project perspectives, a gating fusion factor α is introduced for personalized information fusion. The Moore-Penrose pseudo-inverse analytical method is used to solve the model parameters, complete feature alignment and prediction, and finally generate personalized recommendations.
It enhances the personalized service capabilities and large-scale data adaptation capabilities of the recommendation system, improves recommendation accuracy and feature interaction effects, and enhances the accuracy of model parameter solving.
Smart Images
Figure CN122309858A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of recommendation technology, and in particular to an improved width collaborative filtering method based on bilateral feature alignment. Background Technology
[0002] Personalized recommendation systems have become crucial tools for enterprises to connect users and products and enhance user experience in the digital age. Collaborative filtering, as one of the most widely used technologies in recommendation systems, focuses on analyzing user-item interactions to uncover preferences and provide accurate recommendations to target users. This effectively reduces the cost of acquiring user information and improves platform user retention. Among these, wide-based collaborative filtering recommendation models, based on wide-learning systems, have demonstrated outstanding performance in large-scale and sparse data processing due to their efficient training mechanism and good scalability. Research and application demand for related technologies continues to grow, showcasing significant application potential. However, with the increasing complexity of recommendation scenarios, existing wide-based collaborative filtering methods still have significant shortcomings, making it difficult to achieve ideal recommendation results. Existing research generally acknowledges that user rating criteria naturally differ, while wide-based collaborative filtering has strict requirements for the consistency of feature distribution when constructing input features. This difference in rating criteria directly leads to inconsistent distribution of rating features, thereby interfering with feature interactions and reducing the accuracy of model parameter solving. This is the core flaw of existing technologies.
[0003] Currently, although some studies have optimized wide collaborative filtering in an attempt to alleviate some of the above problems, these optimization methods still have key gaps. They have failed to fundamentally solve the core problem of inconsistent distribution of scoring features, have not designed targeted bias correction strategies in combination with the characteristics of wide learning systems, and have not taken into account both scoring bias correction and model training efficiency, making it difficult to fully leverage the application advantages of wide collaborative filtering in large-scale data scenarios.
[0004] The inventors discovered in their research that existing width-based collaborative filtering methods do not adequately consider the consistency of input feature distribution. The core issue lies in the failure to effectively address the distribution imbalance caused by differences in user rating criteria, and the lack of personalized information extraction and fusion strategies from both the user and project perspectives. This prevents the achievement of feature alignment through reasonable bias correction, ultimately resulting in insufficient recommendation accuracy and generality, and difficulty in adapting to data scenarios with varying sparsity. Therefore, designing an improved width-based collaborative filtering method that can overcome these shortcomings is particularly important. Summary of the Invention
[0005] To address the above problems, this invention proposes an improved width collaborative filtering method based on bilateral feature alignment, comprising:
[0006] The data acquisition and preparation module is used to acquire and clean user-project rating data to obtain the information required by the system.
[0007] The rating decomposition module is used to decompose the target rating of user-project pairs from two perspectives: user and project. It extracts the rating deviation between the two perspectives to represent the personalized information of the corresponding perspective.
[0008] The personalized information fusion module introduces a gated fusion factor α to adaptively fuse the personalized scoring deviation information extracted from the user side and the project side, resulting in a fusion deviation matrix. .
[0009] The feature construction module is used to construct features from the fusion bias into features that meet the input requirements of the width learning model, and to complete the standardization and alignment of features to ensure the consistency of feature distribution.
[0010] The width model prediction module is used to input the constructed alignment features into the width learning model, solve the model parameters through the Moore-Penrose pseudo-inverse analytical method, and complete the prediction of fusion bias.
[0011] The final rating prediction module adds the fusion bias predicted by the width model to the user-item rating fusion mean to obtain the final predicted rating of the target user u for the unrated item v. .
[0012] The recommendation result acquisition module is used to find the set of the top n items with the largest predicted ratings for the target user u and make personalized recommendations.
[0013] The improved width collaborative filtering method of this invention includes a rating decomposition module, which is used to decompose each target rating r in the user-item rating matrix R from two independent perspectives: the user side and the item side. uv Decomposed into user-side bias b u Project-side deviation b v , where b u b is used to characterize the personalized bias caused by users' own rating preferences. v This is used to characterize the personalized bias caused by differences in the project's inherent attributes, thereby extracting personalized information from both sides of the perspective. The formula is as follows:
[0014] ,in For user bias, The average rating of the target users. For project deviations, The average rating of the target project.
[0015] The improved width collaborative filtering method of this invention, wherein the personalized information fusion module is used to fuse personalized information from both sides of the viewpoint through a gated fusion factor α to obtain a weighted combination expression for the target score, as shown in the following formula:
[0016] ,in Used to balance the weight of personalized information on the user side and the project side. The fused rating bias term will serve as the prediction target for the subsequent width learning model, aiming to achieve collaborative alignment of bilateral features and improve recommendation accuracy. The fused mean is used for the final rating forecast.
[0017] Integrated rating bias items Its matrix form is represented as The formula is as follows:
[0018]
[0019] ,in and It is a column vector consisting of 1s, and subtraction is performed only at non-missing positions.
[0020] The improved width collaborative filtering method of this invention includes a feature construction module used to obtain input features of the width learning network. Specifically, it includes:
[0021] (1) User-side feature construction module, used to obtain user-side features. The input features are constructed using collaborative filtering information from the user side, as shown in the following formula:
[0022] ,in Given any user-item pair The set of d nearest neighbor users; for The first in One user; Indicates user For the project Fusion bias (Note: 0 is recorded when there is no value); for Zhongyu The similarity is no less than that of User set.
[0023] (2) Project-side feature construction module, used to obtain project-side features. The input features are constructed using collaborative filtering information from the project side, as shown in the following formula:
[0024] ,in Given any user-item pair of The set of nearest neighbor items; for The first in One project; Indicates user For the project Fusion bias (Note: 0 is recorded when there is no value); for Zhongyu The similarity is no less than A collection of projects.
[0025] (3) Feature concatenation module, used to obtain the overall input features. This module combines user-side features... Features of the project The final input features are obtained by concatenating them. .
[0026] The improved width collaborative filtering method of this invention includes a width model prediction module, which uses constructed features as input to a width neural network for training to predict the fusion bias value of non-interacting items. Specifically, this includes:
[0027] (1) Input layer: Input all user-item pairs in the training set The constructed input feature matrix X and output matrix Y:
[0028] ,in Represents any user-item pair in the training set. The constructed feature vector, Represents any user-item pair in the training set. The corresponding fusion deviation, This represents the training set.
[0029] (2) Feature mapping layer, which uses an activation function to map the input feature matrix Mapped to the feature matrix of the mapping layer :
[0030] ,in It is a weight matrix randomly generated from a certain probability distribution. The bias matrix, To increase the dimensionality of the feature set, For the first One ReLU activation function. All row vectors in the array are identical, and each row vector is denoted as . .
[0031] Mapped feature matrix of the mapping layer as follows:
[0032]
[0033] (3) Feature enhancement layer, which uses activation functions to enhance the feature matrix. Mapped to the feature matrix of the enhancement layer :
[0034] ,in It is a weight matrix randomly generated from a certain probability distribution. The bias matrix, To increase the dimensionality of the feature set, For the first A Tanh activation function. All row vectors in the array are identical, and each row vector is denoted as . .
[0035] Enhanced feature matrix of the enhancement layer as follows:
[0036]
[0037] The weight W is calculated as follows:
[0038] ,in It is the identity matrix. It's a hyperparameter.
[0039] (4) Output layer, obtain any user-project pair The prediction result of the fusion bias is given by the following formula:
[0040] ,in This indicates that the user project will be... Corresponding feature vector The resulting feature vector of the mapping layer is input into the trained BLS network. For the corresponding enhancement layer feature vector, This indicates the predicted fusion bias.
[0041] The improved width collaborative filtering recommendation method of this invention includes a final rating prediction module, which adds the fusion bias predicted by the width model to the user-item rating fusion mean to obtain the final predicted rating of the target user u for the unrated item v. The calculation formula is as follows:
[0042] ,in For the corresponding predicted fusion bias, The average rating of the target users. The average rating of the target project.
[0043] The improved width collaborative filtering method of this invention includes a recommendation result acquisition module for acquiring a recommendation list for a target user u. This includes setting the number of items n recommended by the system for the target user, acquiring the predicted rating set of the target user u on unrated items from the rating prediction acquisition module, sorting the predicted rating values in the set from largest to smallest, and selecting the top n items with the largest predicted values to recommend to the target user u, thereby forming a personalized recommendation list for the target user u.
[0044] The improved width collaborative filtering method described in this invention extracts personalized bias information from both the user and project perspectives. At the same time, it introduces a gating fusion factor to achieve feature distribution alignment and efficient fusion. This has a very positive effect on improving the width collaborative filtering model in optimizing feature interaction effects and parameter solution accuracy, thereby further enhancing the personalized service capability and large-scale data adaptation capability of the recommendation system. Attached Figure Description
[0045] Figure 1 This is a schematic diagram of the method process of the present invention;
[0046] Figure 2 This is a system framework diagram of the present invention;
[0047] Figure 3 This is a comparison of the HR@5 and NDCG@5 values of this invention with other methods across three datasets. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description, in conjunction with the accompanying drawings, provides an improved width collaborative filtering method based on bilateral feature alignment. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit it. Any changes, modifications, additions, alterations, or substitutions made by those skilled in the art within the scope of this invention should be covered by the claims of this invention.
[0049] Figure 1 This is a schematic diagram illustrating the process of the improved width collaborative filtering method based on bilateral feature alignment according to the present invention. From... Figure 1As can be seen, the improved width collaborative filtering method proposed in this invention includes: obtaining and processing user-item rating data from a database to obtain a standardized rating matrix; decomposing the target rating from both user and item perspectives, extracting user-side rating bias and item-side rating bias to represent personalized information from both perspectives; introducing a gating fusion factor to weightedly fuse the personalized rating biases from the user and item sides to obtain fusion bias features; constructing features that meet the input requirements of the width learning model and completing feature alignment; inputting the aligned features into the width learning model, solving the model parameters using the Moore-Penrose pseudo-inverse analytical method to predict the fusion bias; adding the fusion bias predicted by the width model to the mean of the user-item rating fusion to calculate the final predicted rating of the target user on unrated items; sorting the predicted ratings of all unrated items, selecting the top n items with the highest predicted ratings, and generating and pushing personalized recommendation results to the target user.
[0050] Figure 2 This is a framework diagram of the improved width collaborative filtering system based on bilateral feature alignment according to the present invention. From Figure 2As can be seen, the improved width collaborative filtering method and system proposed in this invention includes a data acquisition and preprocessing module, a rating decomposition module, a personalized information fusion module, a feature construction and alignment module, a width model prediction module, a final rating calculation module, and a recommendation result generation module. The data acquisition and preprocessing module extracts user-item rating data from the database and processes it into a standardized rating matrix format required by subsequent modules. The rating decomposition module decomposes the target rating from both the user and item perspectives based on the standardized data provided by the data acquisition and preprocessing module, extracting user-side rating deviation and item-side rating deviation respectively to represent personalized information from both perspectives. The personalized information fusion module introduces a gating fusion factor to weight and fuse the personalized deviations from the user and item sides based on the bilateral deviation information obtained by the rating decomposition module, obtaining a fusion deviation feature. The feature construction and alignment module calculates the fusion deviation feature obtained by the personalized information fusion module based on the fusion deviation feature obtained by the personalized information fusion module. The feature distribution is aligned with the input requirements of the width learning model. The width model prediction module constructs aligned features with those provided by the alignment module and inputs them into the width learning model. The Moore-Penrose pseudo-inverse analytical method is used to solve for the model parameters and predict the fusion bias. The final score calculation module adds the fusion bias prediction result obtained by the width model prediction module to the user-item score fusion mean to calculate the final predicted score of the target user on the unrated items. The recommendation result generation module sorts the predicted score set obtained by the final score calculation module and recommends the top n items with the highest predicted scores to the target user.
[0051] Furthermore, the following example illustrates this further:
[0052] Suppose there are M users U = {u1, u2, ..., u...} M} and N items I = {i1, i2, ..., i N The user-item rating matrix is represented by [R]. M×N Indicates; r ui This represents the user's rating for item i.
[0053] 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 similarity recommendation system are as follows:
[0054] S1: The data acquisition and preparation module retrieves the required data information from the database, including user ID, project ID, and rating information;
[0055] S2: The rating decomposition module converts user rating information into deviation information between the user side and the project side;
[0056] From two independent perspectives, the user-project rating matrix R and the target rating r are analyzed. uv Decomposed into user-side bias b u Project-side deviation b v , where b u b is used to characterize the personalized bias caused by users' own rating preferences. v This is used to characterize the personalized bias caused by differences in the project's inherent attributes, thereby extracting personalized information from both sides of the perspective. The expression is as follows:
[0057]
[0058] in, For user bias, The average rating of the target users. For project deviations, The average rating of the target project.
[0059] S3: The personalized information fusion module fuses personalized information from both sides of the perspective by introducing a gated fusion factor α, resulting in a weighted combination expression for the target score, as shown in the following formula:
[0060] ,
[0061] Among them, China Used to balance the weight of personalized information on the user side and the project side. The fused rating bias term will serve as the prediction target for the subsequent width learning model, aiming to achieve collaborative alignment of bilateral features and improve recommendation accuracy. The fused mean is used for the final rating forecast.
[0062] Integrated rating bias items Its matrix form is represented as The formula is as follows:
[0063]
[0064] ,in and It is a column vector consisting of 1s, and subtraction is performed only at non-missing positions.
[0065] S4: The feature construction module constructs a width learning network based on the fusion bias information obtained in S3 to input features. Features are constructed from both the user and project perspectives.
[0066] The user-side feature construction module is used to obtain user-side features. The input features are constructed using collaborative filtering information from the user side, as shown in the following formula:
[0067]
[0068] in, Given any user-item pair The set of d nearest neighbor users; for The first in One user; Indicates user For the project Fusion bias (Note: 0 is recorded when there is no value); for Zhongyu The similarity is no less than The user set.
[0069] The project-side feature construction module is used to obtain project-side features. The input features are constructed using collaborative filtering information from the project side, as shown in the following formula:
[0070]
[0071] in, Given any user-item pair of The set of nearest neighbor items; for The first in One project; Indicates user For the project Fusion bias (Note: 0 is recorded when there is no value); for Zhongyu The similarity is no less than A collection of projects.
[0072] Then, user-side features Features of the project The final input features are obtained by concatenating them. .
[0073] S5: Width Model Prediction Module, used to train a width neural network by inputting the constructed features to predict the fusion bias value of non-interacting items, specifically including:
[0074] (1) Input layer: Input all user-item pairs in the training set The constructed input feature matrix X and output matrix Y:
[0075]
[0076] in, Represents any user-item pair in the training set. The constructed feature vector, Represents any user-item pair in the training set. The corresponding fusion deviation, This represents the training set.
[0077] (2) Feature mapping layer, which uses an activation function to map the input feature matrix Mapped to the feature matrix of the mapping layer :
[0078]
[0079] in, It is a weight matrix randomly generated from a certain probability distribution. The bias matrix, To increase the dimensionality of the feature set, For the first One ReLU activation function. All row vectors in the array are identical, and each row vector is denoted as . .
[0080] Mapped feature matrix of the mapping layer as follows:
[0081]
[0082] (3) Feature enhancement layer, which uses activation functions to enhance the feature matrix. Mapped to the feature matrix of the enhancement layer :
[0083]
[0084] in, It is a weight matrix randomly generated from a certain probability distribution. The bias matrix, To increase the dimensionality of the feature set, For the first A Tanh activation function. All row vectors in the array are identical, and each row vector is denoted as . .
[0085] Enhanced feature matrix of the enhancement layer as follows:
[0086]
[0087] The weight W is calculated as follows:
[0088]
[0089] in, It is the identity matrix. It's a hyperparameter.
[0090] (4) Output layer, obtain any user-project pair The prediction result of the fusion bias is given by the following formula:
[0091]
[0092] in, This indicates that the user project will be... Corresponding feature vector The resulting feature vector of the mapping layer is input into the trained BLS network. For the corresponding enhancement layer feature vector, This indicates the predicted fusion bias.
[0093] S6: The final score prediction module effectively integrates the fusion mean obtained in S3 and the prediction fusion bias obtained in S5 to obtain... ;
[0094] The fusion bias predicted by the width model is added to the mean of the user-item rating fusion to obtain the final predicted rating of the target user u for the unrated item v. The calculation formula is as follows:
[0095]
[0096] in, For the corresponding predicted fusion bias, The average rating of the target users. The average rating of the target project.
[0097] S7: The recommendation result acquisition module, based on the set of predicted rating items calculated in S6 and the number of recommended items n set by the system, first sorts the rating values in the predicted set from largest to smallest, and then selects the top n items with the largest predicted values to recommend to the target user u, so as to form a personalized recommendation list for the target user u.
[0098] Figure 3 The proposed method was measured in two metrics: hit rate (HR) and standardized discount cumulative return (NDCG). Their calculation methods and measurement details are shown below:
[0099] Hit Rate (HR): Used to measure whether the actual recommended items in the test set appear on the top-k predicted recommendation list. It indicates the model's item recommendation ability and is calculated as follows:
[0100]
[0101] Where m represents the number of users in the recommendation system, and hits(i) represents the proportion of the first k predicted recommendation items of the i-th user in the actual recommendation item set.
[0102] Standardized Discount Cumulative Return (NDCG): This indicates the quality of the model's item ranking recommendations by assigning higher scores to higher-ranking clicks. The calculation formula is as follows:
[0103]
[0104]
[0105]
[0106] Where DCG@k and IDCG@k represent the discounted cumulative gain and the ideal DCG, respectively, rel p The `rel` parameter is used to display the recommendation relevance of the item at position `p`, i.e., if the predicted recommended item appears in the actual recommendation list. p =1 otherwise =0. It's worth noting that we assume a project is recommended if its predicted or actual rating exceeds the median of the rating range.
Claims
1. An improved width collaborative filtering method based on bilateral feature alignment, characterized in that, include: The data acquisition and preparation module is used to acquire and clean user-project rating data to obtain the information required by the system. The rating decomposition module is used to decompose the target rating of user-project pairs from two perspectives: user and project. It extracts the rating deviation between the two perspectives to represent the personalized information of the corresponding perspective. The personalized information fusion module is used to introduce a gating fusion factor and adaptively fuse the personalized scoring deviation information extracted from the user side and the project side to obtain a fusion deviation matrix. The feature construction module is used to construct features that meet the input requirements of the width learning model from the fusion bias, and to complete the standardization and alignment of features to ensure the consistency of feature distribution. The width model prediction module is used to input the constructed alignment features into the width learning model, solve the model parameters through the Moore-Penrose pseudo-inverse analytical method, and complete the prediction of fusion bias. The final rating prediction module is used to add the fusion bias predicted by the width model to the user-item rating fusion mean to obtain the final predicted rating of the target user for the unrated item; The recommendation results acquisition module is used to find the set of the top n items with the highest predicted ratings for personalized recommendations to the target user.
2. The improved width collaborative filtering method based on bilateral feature alignment as described in claim 1, characterized in that, The rating decomposition module is used to decompose each target rating r in the user-project rating matrix R from two independent perspectives: the user side and the project side. u,v Decomposed into user-side bias b u Project-side deviation b v , where b u b is used to characterize the personalized bias caused by users' own rating preferences. v This is used to characterize the personalized bias caused by differences in the project's inherent attributes, thereby extracting personalized information from both sides of the perspective. The formula is as follows: ,in For user bias, The average rating of the target users. For project deviations, This represents the average rating of the target project.
3. The improved width collaborative filtering method based on bilateral feature alignment as described in claim 1, characterized in that, The personalized information fusion module is used to fuse personalized information from both sides of the viewpoint through a gated fusion factor α to obtain a weighted combination expression for the target score, as shown in the following formula: ,in It is used to balance the weight of personalized information on the user side and the project side. This is the fused rating bias term, which will serve as both input and prediction target for the subsequent width-learning network. The average of the merged scores will be used for the final rating prediction; Integrated rating bias items Its matrix form is represented as The formula is as follows: ， ,in and It is a column vector consisting of 1s, and subtraction is performed only at non-missing positions.
4. The improved width collaborative filtering method based on bilateral feature alignment as described in claim 1, characterized in that, The feature construction module is used to obtain the input features of the width learning network. Specifically, it includes: (1) User-side feature construction module, used to obtain user-side features. The input features are constructed using collaborative filtering information from the user side, as shown in the following formula: ,in Given any user-item pair (u,v), the set of d nearest neighbor users; for The first in One user; Indicates user For the project Fusion bias (Note: 0 is recorded when there is no value); for Zhongyu The similarity is no less than The user set; (2) Project-side feature construction module, used to obtain project-side features. The input features are constructed using collaborative filtering information from the project side, as shown in the following formula: ,in For any user-item pair (u,v) The set of nearest neighbor items; for The first in One project; Indicates user For the project Fusion bias (Note: 0 is recorded when there is no value); for Zhongyu The similarity is no less than A collection of projects; (3) Feature concatenation module, used to obtain the overall input features and combine the user-side features Features of the project The final input features are obtained by concatenating them. , This indicates the concatenation of vectors.
5. The improved width collaborative filtering method based on bilateral feature alignment as described in claim 1, characterized in that, The width model prediction module is used to train a width neural network by inputting the constructed features into it, in order to predict the fusion bias value of non-interactive items, specifically including: (1) Input layer: Input all user-item pairs in the training set The constructed input feature matrix X and output matrix Y: ,in Represents any user-item pair in the training set. The constructed feature vector, Represents any user-item pair in the training set. The corresponding fusion deviation, Represents the training set, Indicates the number of training sets; (2) Feature mapping layer, which uses an activation function to map the input feature matrix Mapped to the feature matrix of the mapping layer : Where N is the number of mapping feature groups, It is a weight matrix randomly generated by the feature mapping layer. This is the bias matrix of the feature mapping layer. Let be the dimension of the mapping feature set. For the first One ReLU activation function, All row vectors in the array are identical, and each row vector is denoted as . The feature matrix of the mapped layer is: , Indicates the concatenation of vectors; (3) Feature enhancement layer, which uses activation functions to enhance the feature matrix. Mapped to the feature matrix of the enhancement layer : Where M is the number of enhanced feature groups, It is a weight matrix randomly generated by the feature enhancement layer. Here is the bias matrix for the feature enhancement layer. To increase the dimensionality of the feature set, For the first One Tanh activation function, All row vectors in the array are identical, and each row vector is denoted as . The enhanced feature matrix of the enhancement layer is: ; The weight W is calculated as follows: ,in It is the identity matrix. It's a hyperparameter; (4) Output layer, obtain any user-project pair The prediction result of the fusion bias is given by the following formula: ,in This indicates that the user project will be... Corresponding feature vector The resulting feature vector of the mapping layer is input into the trained BLS network. For the corresponding enhancement layer feature vector, This indicates the predicted fusion bias.
6. The improved width collaborative filtering method based on bilateral feature alignment as described in claim 1, characterized in that, The final rating prediction module is used to add the fusion bias predicted by the width model to the user-item rating fusion mean to obtain the final predicted rating of the target user u for the unrated item v. The calculation formula is as follows: ,in For the corresponding predicted fusion bias, The average rating of the target users. This represents the average rating of the target project.
7. The improved width collaborative filtering method based on bilateral feature alignment as described in claim 1, characterized in that, The recommendation result acquisition module is used to acquire a recommendation list for the target user u. This includes setting the number of items n recommended by the system to the target user, acquiring the predicted rating set of the target user u on unrated items from the rating prediction acquisition module, sorting the predicted rating values in the set from largest to smallest, and selecting the top n items with the largest predicted values to recommend to the target user u, so as to form a personalized recommendation list for the target user u.