A recommendation algorithm based on a label neighborhood model and a label matrix factorization model

Abstract

The application discloses a kind of recommendation algorithm based on label neighborhood model and label matrix decomposition model, it is related to user recommendation technical field, including the construction UTagUser-CF algorithm and TagMF algorithm, according to the UTagUser-CF algorithm and TagMF algorithm constructed to construct UTagUser-TagMF algorithm;The recommendation algorithm based on label neighborhood model and label matrix decomposition model, by using TagMF algorithm, so that the algorithm has lower recommendation error, better recommendation effect, using label information model helps matrix decomposition model to increase implicit feature information, to further solve the cold start problem and data sparsity problem, to solve the accuracy problem, using UTagUser-TagMF algorithm when facing the cold start problem, so that the algorithm has better recommendation performance, can effectively solve the cold start problem and data sparsity problem.

Description

Technical Field

[0001] This invention relates to the field of user recommendation technology, specifically a recommendation algorithm based on a label neighborhood model and a label matrix factorization model. Background Technology

[0002] In the information age, network data is growing rapidly. While this data promotes the overall development of the internet, it also causes information overload for most users. Faced with massive amounts of data, users struggle to extract valuable information and find what they need. Recommendation systems can help users extract information of interest from massive amounts of data, thus solving the problem of information overload. Recommendation algorithms are the foundation of recommendation systems. These algorithms integrate data such as user preferences, rating behavior, item tags, and domain environment information to analyze the relevance between users and items, identify the items users like or need most, and provide recommendation services. Recommendation algorithms are generally divided into three categories: content-based recommendation algorithms, collaborative filtering (CF) algorithms, and hybrid recommendation algorithms. Collaborative filtering is the most representative recommendation algorithm. This algorithm does not require collecting additional information data; it can identify user-liked items and make recommendations based solely on rating data. Although significant progress has been made in the theoretical research and practical applications of collaborative filtering algorithms, several problems still exist, including cold start, data sparsity, scalability, and accuracy issues. These problems can lead to increased recommendation errors, deterioration in recommendation quality, decreased algorithm performance, and inaccurate predictions in recommendation systems, thus affecting users' access to information. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a recommendation algorithm based on a label neighborhood model and a label matrix factorization model, which solves the problems mentioned in the background.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solution: a recommendation algorithm based on a tag neighborhood model and a tag matrix factorization model, including constructing a UTagUser-CF algorithm and a TagMF algorithm, and constructing a UTagUser-TagMF algorithm based on the constructed UTagUser-CF algorithm and TagMF algorithm;

[0005] First, the formula for calculating the user's predicted rating of a project in the UTagUser-CF algorithm is as follows (1);

[0006] (1)

[0007] This indicates the overall similarity between users;

[0008] Next is the formula for calculating the user's predicted rating of the project in the TagMF algorithm, as shown in equation (2).

[0009] (2)

[0010] The UTagUser-TagMF algorithm uses a combination of weight coefficients. The two algorithms are combined, and the formula for calculating the predicted score is shown in equation (3).

[0011] (3)

[0012] In equation (3), Represents the predicted rating of the UTagUser-CF algorithm. Represents the predicted score of the TagMF algorithm;

[0013] Optionally, the UTagUser-CF algorithm is constructed as follows:

[0014] The calculation steps for establishing a label model are as follows:

[0015] S1: Model the target tags for the project and construct the project tag model matrix. Assume there are a certain number of project tag types. One, to establish for each project Dimensional vector, item The label vector is Assuming the project consists of tag type 1 and tag type 4, then the first and fourth dimensions of the project tag vector are 1, and the rest are 0. Therefore, the project tag vector... The value is shown in equation (1-1).

[0016] (1-1)

[0017] By constructing a tag vector for each project, a project tag matrix (ITag) can be generated. The size of this matrix is... Where m is the number of projects. It represents the number of tag types, and the matrix ITag is in the form of equation (1-2).

[0018] (1-2)

[0019] S2: Build a user tag model based on the project tag model and rating matrix, generating a tag for each user. The user's label vector is a dimensional vector. The user's calculation is performed using the following formula. The label vector represents the user's rating for each label type. The calculation formula is shown in equation (1-3).

[0020] (1-3)

[0021] In equation (1-3), This represents the label vector of user u. This represents the label model for item i. The set of items representing user u's rating. A tag vector is constructed for each user, generating a user tag matrix UTag, the size of which is... Where n is the number of users, It represents the number of tag types, and the matrix UTag is in the form of equation (1-4).

[0022] (1-4)

[0023] The tag information model is integrated into the Neighborhood-CF algorithm to construct the UTagUser-CF algorithm;

[0024] Based on the user tag model matrix UTag, each user has their own tag vector. For the user tag vector matrix, the widely used Pearson coefficient can be used to calculate the similarity between user tag vectors, that is, to assess the similarity of users' preferences for tag types. The formula for calculating user tag vector similarity is shown in equation (1-5).

[0025] (1-5)

[0026] In equation (1-5), Indicates user and users Similarity in user tagging models This represents the label vector of user u. This represents the average value of the user's tag vector. This represents the value of the i-th dimension of the label vector for user u;

[0027] After obtaining the user tag similarity, an activation function is used to process the user tag similarity matrix. Similarity values ​​less than a set threshold L are removed from the user tag similarity matrix; the activation function is shown in equation (1-6).

[0028] (1-6)

[0029] Using the UTag matrix as a basis, and using Equation (1-5) to calculate the similarity of tag models between users, and then using Equation (1-6) to preprocess the tag similarity matrix, the resulting tag similarity matrix is ​​used as the similarity matrix. This algorithm is called the User Tag Model-Based Collaborative Filtering (UTag-CF) algorithm. This algorithm uses tag information as data and can be used as a content-based recommendation algorithm.

[0030] There are two different levels of user similarity: one based on the rating matrix and the other based on the user tag model matrix. Combining these two different levels of user similarity, the formula for combining the two similarity matrices is shown in equation (1-7):

[0031] (1-7)

[0032] In equation (1-7), This indicates the overall similarity between users. Indicates the similarity of user ratings. Indicates the similarity of user tag models. This represents the weighting coefficient between the two, and subsequent experimental designs will consider the impact of the weighting coefficient on the algorithm.

[0033] The comprehensive similarity matrix serves as the result of the first stage of similarity calculation in the collaborative filtering algorithm. The second stage neighbor selection method uses either the number of neighbor sets or the comprehensive similarity threshold. The formula for calculating the prediction score in the third stage is shown in equation (1).

[0034] (1)

[0035] Optionally, the construction of the TagMF algorithm:

[0036] In the tag information model, there are user tag models and item tag models. Assuming the total number of tag types in the item table is T, then the user tag model matrix is... The project tag model matrix is .

[0037] By combining these two label model matrices and adding them to the original two initialization matrices, the two latent feature matrices contain more contextual information. The sizes of the two initialization matrices must be respectively... and The calculation formulas are shown in equations (2-1) and (2-2).

[0038] (2-1)

[0039] (2-2)

[0040] After completing the calculations of equations (2-1) and (2-2), two latent feature matrices with label context information are obtained. These two matrices are then standardized using the min-max normalization method, as shown in equation (2-3):

[0041] (2-3)

[0042] In equation (2-3), Represents the minimum value of the matrix. Represents the maximum value of the matrix. This represents each value in the matrix;

[0043] After completing the initialization in the first stage, the second stage is consistent with the commonly used MF algorithm, employing stochastic gradient descent to optimize the loss function while continuously iterating and training to obtain the matrix of the optimal solution. and ;

[0044] Therefore, the TagMF algorithm is obtained as shown in equation (2):

[0045] (2)

[0046] Optionally, after constructing the UTagUser-TagMF algorithm, it needs to be validated. The validation content includes the impact of different learning step sizes on the algorithm; the impact of different regularization coefficients on the algorithm; the impact of different combination weight coefficients on the algorithm; the impact of different dataset sizes on the algorithm; the impact of the cold start problem on the algorithm; and the comparison with other algorithms.

[0047] This invention provides a recommendation algorithm based on a label neighborhood model and a label matrix factorization model, which has the following advantages:

[0048] 1. This recommendation algorithm based on the label neighborhood model and label matrix factorization model achieves lower recommendation error and better recommendation performance by employing the TagMF algorithm. The use of the label information model helps to increase the implicit feature information of the matrix factorization model, thereby further solving the cold start problem and the data sparsity problem, and thus solving the accuracy problem. The use of the UTagUser-TagMF algorithm makes this algorithm perform better in the face of the cold start problem, and can effectively solve the cold start problem and the data sparsity problem.

[0049] 2. This recommendation algorithm, based on a label neighborhood model and a label matrix factorization model, constructs the TagMF algorithm by integrating the label information model into the MF algorithm. This algorithm incorporates the user label model and the item label model into the two initialization matrices in the first stage, addressing the problem of insufficient implicit feature information in the initialization matrices. This allows for the discovery of more implicit feature correlations between users and items, effectively improving the algorithm's recommendation quality. By effectively combining the neighborhood model and the matrix factorization model, this algorithm can discover correlations through strong correlations between users and also through the correlation of users' overall ratings of items. It possesses the advantages of both models while complementing each other's shortcomings. Attached Figure Description

[0050] Figure 1 This is a graph showing the mean absolute error of the algorithm with different learning step sizes in this invention;

[0051] Figure 2 This is a graph showing the absolute error of the algorithm with different regularization coefficients in this invention.

[0052] Figure 3 This is a graph showing the mean absolute error of the MF algorithm under different weighting coefficients in this invention.

[0053] Figure 4 This is a graph showing the mean absolute error of the MF algorithm under different datasets in this invention.

[0054] Figure 5 This diagram shows a comparison of three collaborative filtering algorithms based on matrix factorization according to the present invention. Detailed Implementation

[0055] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0056] This invention provides a technical solution: a recommendation algorithm based on a tag neighborhood model and a tag matrix factorization model, characterized in that: it includes constructing a UTagUser-CF algorithm and a TagMF algorithm, and constructing a UTagUser-TagMF algorithm based on the constructed UTagUser-CF algorithm and TagMF algorithm;

[0057] First, the formula for calculating the user's predicted rating of a project in the UTagUser-CF algorithm is as follows (1);

[0058] (1)

[0059] This indicates the overall similarity between users;

[0060] Next is the formula for calculating the user's predicted rating of the project in the TagMF algorithm, as shown in equation (2).

[0061] (2)

[0062] The UTagUser-TagMF algorithm uses a combination of weight coefficients. The two algorithms are combined, and the formula for calculating the predicted score is shown in equation (3).

[0063] (3)

[0064] In equation (3), Represents the predicted rating of the UTagUser-CF algorithm. Represents the predicted score of the TagMF algorithm;

[0065] Among them: Construction of the UTagUser-CF algorithm:

[0066] The calculation steps for establishing a label model are as follows:

[0067] S1: Model the target tags for the project and construct the project tag model matrix. Assume there are a certain number of project tag types. One, to establish for each project Dimensional vector, item The label vector is Assuming the project consists of tag type 1 and tag type 4, then the first and fourth dimensions of the project tag vector are 1, and the rest are 0. Therefore, the project tag vector... The value is shown in equation (1-1).

[0068] (1-1)

[0069] By constructing a tag vector for each project, a project tag matrix (ITag) can be generated. The size of this matrix is... Where m is the number of projects. It represents the number of tag types, and the matrix ITag is in the form of equation (1-2).

[0070] (1-2)

[0071] S2: Build a user tag model based on the project tag model and rating matrix, generating a tag for each user. The user's label vector is a dimensional vector. The user's calculation is performed using the following formula. The label vector represents the user's rating for each label type. The calculation formula is shown in equation (1-3).

[0072] (1-3)

[0073] In equation (1-3), This represents the label vector of user u. This represents the label model for item i. The set of items representing user u's rating. A tag vector is constructed for each user, generating a user tag matrix UTag, the size of which is... Where n is the number of users, It represents the number of tag types, and the matrix UTag is in the form of equation (1-4).

[0074] (1-4)

[0075] Tag information models can be used as contextual information to improve the Neighborhood-CF algorithm. Project tag models can be used to solve the cold start problem for new projects, and user tag models can be used to solve the data sparsity problem caused by insufficient user ratings.

[0076] The tag information model is integrated into the Neighborhood-CF algorithm to construct the UTagUser-CF algorithm;

[0077] Based on the user tag model matrix UTag, each user has their own tag vector. For the user tag vector matrix, the widely used Pearson coefficient can be used to calculate the similarity between user tag vectors, that is, to assess the similarity of users' preferences for tag types. The formula for calculating user tag vector similarity is shown in equation (1-5).

[0078] (1-5)

[0079] In equation (1-5), Indicates user and users Similarity in user tagging models This represents the label vector of user u. This represents the average value of the user's tag vector. This represents the value of the i-th dimension of the label vector for user u;

[0080] After obtaining the user tag similarity, an activation function is used to process the user tag similarity matrix. Similarity values ​​less than a set threshold L are removed from the user tag similarity matrix. The activation function is shown in equation (1-6):

[0081] (1-6)

[0082] Using the UTag matrix as a basis, and using Equation (1-5) to calculate the similarity of tag models between users, and then using Equation (1-6) to preprocess the tag similarity matrix, the resulting tag similarity matrix is ​​used as the similarity matrix. This algorithm is called the User Tag Model-Based Collaborative Filtering (UTag-CF) algorithm. This algorithm uses tag information as data and can be used as a content-based recommendation algorithm.

[0083] There are two different levels of user similarity: one based on the rating matrix and the other based on the user tag model matrix. Combining these two different user similarity metrics, tag model similarity plays a supporting role as contextual information, integrating tag information into the rating similarity. High tag model similarity promotes rating similarity, while low tag model similarity inhibits it. Based on the combined similarity matrix, when users search for neighbors, they can find users with similar rating behaviors and similar preferences for item tag types. The formula for combining the two similarity matrices is shown in equation (1-7):

[0084] (1-7)

[0085] In equation (1-7), This indicates the overall similarity between users. Indicates the similarity of user ratings. Indicates the similarity of user tag models. This represents the weighting coefficient between the two, and subsequent experimental designs will consider the impact of the weighting coefficient on the algorithm.

[0086] The comprehensive similarity matrix serves as the result of the first stage of similarity calculation in the collaborative filtering algorithm. The second stage neighbor selection method uses either the number of neighbor sets or the comprehensive similarity threshold. The formula for calculating the prediction score in the third stage is shown in equation (1).

[0087] (1)

[0088] The construction of the TagMF algorithm is as follows:

[0089] In the first stage of the MF algorithm, two matrices need to be initialized. However, during matrix initialization, the initial values ​​of the two matrices are random numbers in the range of 0 to 1, resulting in insufficient latent feature information and severely impacting recommendation performance. To address this, a tag information model is integrated into the initialization process of the MF algorithm, improving the matrix factorization model and constructing a TagMF algorithm. This allows the initialized matrices to contain more latent feature information, enabling the discovery of more latent feature relationships between users and items.

[0090] In the tag information model, there are user tag models and item tag models. Assuming the total number of tag types in the item table is T, then the user tag model matrix is... The project tag model matrix is .

[0091] By combining these two label model matrices and adding them to the original two initialization matrices, the two latent feature matrices contain more contextual information. The sizes of the two initialization matrices must be respectively... and The calculation formulas are shown in equations (2-1) and (2-2).

[0092] (2-1)

[0093] (2-2)

[0094] After completing the calculations of equations (2-1) and (2-2), two latent feature matrices with label context information are obtained. Because the dimension of the label information model is inconsistent with the scale of the latent feature information in the initialization matrix, these two matrices are standardized. Standardization eliminates the influence of differences in scale. The min-max standardization method is used, and the calculation formula is shown in equation (2-3).

[0095] (2-3)

[0096] In equation (2-3), Represents the minimum value of the matrix. Represents the maximum value of the matrix. This represents each value in the matrix;

[0097] After completing the initialization in the first stage, the second stage is consistent with the commonly used MF algorithm, employing stochastic gradient descent to optimize the loss function while continuously iterating and training to obtain the matrix of the optimal solution. and ;

[0098] Therefore, the TagMF algorithm is obtained as shown in equation (2):

[0099] (2)

[0100] Example

[0101] like Figure 1As shown: After constructing the UTagUser-TagMF algorithm, it needs to be validated. The validation content includes the impact of different learning step sizes on the algorithm; the impact of different regularization coefficients on the algorithm; the impact of different combination weight coefficients on the algorithm; the impact of different dataset sizes on the algorithm; the impact of the cold start problem on the algorithm; and comparison with other algorithms.

[0102] The impact of different learning step sizes on the algorithm:

[0103] like Figure 1 The figure shows the MAE experimental data for the MF algorithm, TagMF algorithm, and UTagUser-TagMF algorithm under different learning step sizes. The iteration number is 20, and the regularization coefficient is... The value is 0.02, where the combined weight coefficient of the UTagUser-TagMF algorithm is... It is 0.1.

[0104] observe Figure 1 The data reveals that the recommendation error (MAE) of the MF, tagMF, and UTagUser-TagMF algorithms first decreases and then increases as the learning step size increases from 0.00002 to 0.01. When the learning step size is 0.02, the recommendation error of all three algorithms increases significantly compared to before.

[0105] pass Figure 1 Based on observation and analysis, the following conclusions can be drawn from this experiment:

[0106] (1) When the learning step size is too large, the algorithm oscillates and the vectors of users and items cannot converge, resulting in a high recommendation error and poor recommendation effect.

[0107] (2) When the learning step size is too small, the algorithm cannot converge in time, the convergence speed is slow, and the algorithm recommendation error is high.

[0108] (3) When the learning step size is around 0.002, the three algorithms can maintain good performance.

[0109] (4) Compared with the recommendation error of the three algorithms, under the condition of appropriate learning step size, TagMF algorithm and UTagUser-TagMF algorithm both showed better recommendation effect, indicating that incorporating the tag information model as context information into the recommendation algorithm can bring about improvement, provide more implicit feature information, and thus improve the algorithm performance.

[0110] In summary, learning step size This affects all three algorithms. A learning step size that is too large can cause the algorithm to oscillate, failing to converge and find the optimal value, thus increasing the recommendation error. A learning step size that is too small results in slow convergence and a higher recommendation error. It can also be seen that incorporating the tag information model into the matrix factorization model is a good improvement, reducing the recommendation error and improving the algorithm's recommendation performance.

[0111] The impact of different regularization coefficients on the algorithm:

[0112] Figure 2 The MF algorithm, TagMF algorithm, and UTagUser-TagMF algorithm have different regularization coefficients. The following MAE experimental data is used. The iteration count is 20, and the learning step size is... The value is 0.002, where the combined weight coefficient of the UTagUser-TagMF algorithm is... It is 0.1.

[0113] observe Figure 2 Regularization coefficients can be observed. When the regularization coefficient is 0.0002, 0.002, 0.02, and 0.2, the recommendation error of the three algorithms does not change significantly; When the value is 2, the recommendation error of the three algorithms increases, and the increase is relatively large.

[0114] pass Figure 2 Based on observation and analysis, the following conclusions can be drawn from this experiment:

[0115] (1) When the regularization coefficient is small, the recommendation error of the algorithm remains basically unchanged, which can prevent overfitting while maintaining the recommendation effect of the algorithm.

[0116] (2) When the regularization coefficient is too large, the recommendation error of the algorithm increases and the magnitude is not small, indicating that an excessively large regularization coefficient will affect the recommendation error of the algorithm and affect the optimization of the loss function.

[0117] In summary, the regularization coefficient affects the algorithm. The regularization coefficient should not be too large, otherwise it will affect the algorithm's performance and recommendation effectiveness. Subsequent experiments will maintain a regularization coefficient of 0.02.

[0118] The impact of different combination weight coefficients on the algorithm:

[0119] Figure 3 The UTagUser-TagMF algorithm uses different weighting coefficients. The dataset includes MAE experimental data, as well as experimental data for the MF and TagMF algorithms. All three algorithms have an iteration count of 20 (step) and a learning step size of [missing information]. The regularization coefficient is 0.002. It is 0.02.

[0120] observe Figure 3 The data reveals that as the weighting coefficients are combined... As the number of cases increases, the algorithm's recommendation error continues to grow.

[0121] pass Figure 3 Based on observation and analysis, the following conclusions can be drawn from this experiment:

[0122] (1) TagMF algorithm in terms of iteration number step and learning step size Regularization coefficient Under the same conditions, the recommendation error is lower, indicating that the label neighborhood model does not directly reduce the error of the label matrix decomposition model.

[0123] (2) The MF algorithm in terms of iteration number step and learning step size Regularization coefficient Under the same conditions, the recommendation error is higher, indicating that the UTagUser-TagMF algorithm still has a better recommendation effect than the original MF algorithm when the weight coefficient is small.

[0124] In summary, incorporating weighting coefficients affects the UTagUser-TagMF algorithm. Combining the tag matrix factorization model with the tag neighborhood model enhances the discovery of local associations between users, but the recommendation error increases; however, it still outperforms the MF algorithm in recommendation performance. It is necessary to further incorporate weighting coefficients... Within a relatively small range, the label neighborhood model plays a supporting role.

[0125] The impact of different dataset sizes on the algorithm:

[0126] The following text Figure 4 These are experimental data for the MF algorithm, TagMF algorithm, and UTagUser-TagMF algorithm on different datasets. The number of iterations is 20, and the learning step size is... The regularization coefficient is 0.002. It is 0.02, combined with the weighting coefficient. It is 0.1.

[0127] observe Figure 4 It can be observed that the larger the dataset, the smaller the recommendation error of the algorithm; when the dataset is ml-100k, the TagMF algorithm and the UTagUser-TagMF algorithm perform better than the MF algorithm; when the dataset is ml-1M, the TagMF algorithm performs better than the MF algorithm, while the UTagUser-TagMF algorithm does not.

[0128] pass Figure 4Based on observation and analysis, the following conclusions can be drawn from this experiment:

[0129] (1) When the dataset increases, i.e., when the amount of data is large, the matrix factorization model can achieve lower recommendation error and better recommendation results. This indicates that an increase in the amount of data helps the matrix factorization model to better reflect users' rating behavior of items, making the recommendations more accurate.

[0130] (2) When the dataset is large, combining it with the neighborhood model will affect the recommendation error of its own matrix factorization model, thus affecting the recommendation effect. This indicates that when facing large-scale data, the algorithm should pay more attention to the overall rating situation and only explore the correlation between users or items to a limited extent.

[0131] (3) Regardless of the size of the dataset, the label matrix factorization model consistently exhibits lower recommendation error and better recommendation performance. This indicates that integrating the label information model into the matrix factorization model is a good approach, effectively increasing implicit feature information and improving the algorithm's recommendation performance.

[0132] In summary, dataset size affects the MF (Matrix Factorization) algorithm, TagMF (Tag-MF) algorithm, and UTagUser-TagMF (UTagUser-TagMF) algorithm. With increasing data volume, matrix factorization models benefit, leading to more thorough model training and better algorithm performance. Furthermore, when dealing with large-scale datasets, it's advisable to reduce neighborhood correlation mining and instead utilize the implicit features of the overall rating to improve the recommendation algorithm.

[0133] The impact of the cold start problem on algorithms:

[0134] This experiment processed the ml-100k dataset to create a scenario where new users have no rating data upon joining. 20% of users were randomly selected from the training set as new users, and their rating data was set to 0; this dataset is ml-100k(coldstart-20%). Similarly, 50% of users were randomly selected from the training set as new users, and their rating data was set to 0; this dataset is ml-100k(coldstart-50%). The impact of the algorithm on the cold start dataset will be explored below.

[0135] Table 4-1 shows the experimental data of the three algorithms on a normal dataset and a dataset with the cold start problem. Here, ml-100k (coldstart) represents the dataset with the cold start problem, and the improvement ratio is the improvement ratio of the recommendation error in the cold start environment. The number of iterations is 20, and the learning step size is... The regularization coefficient is 0.002. It is 0.02, combined with the weighting coefficient. It is 0.1.

[0136] Table 1. Mean Absolute Error of Algorithms for Datasets with Cold Start Problems

[0137] ml-100k ml-100k (coldstart-20%) ml-100k (coldstart-50%) MF 0.9671 1.1841 1.7098 TagMF 0.8274 0.8993 1.1036 UTagUser-TagMF 0.9077 0.9300 0.9969

[0138] Observing the data in Table 1, it can be found that the recommendation error of all three algorithms increases when faced with a dataset with more new users. In the dataset ml-100k(coldstart-20%), the recommendation error of the TagMF algorithm is smaller than that of the UTagUser-TagMF algorithm. In the dataset ml-100k(coldstart-50%), the recommendation error of the TagMF algorithm is larger than that of the UTagUser-TagMF algorithm.

[0139] Based on the observations and analysis in Table 1, the following conclusions can be drawn from this experiment:

[0140] (1) All three algorithms have higher recommendation errors when faced with datasets with cold start problems, indicating that the increase in the number of new users will affect the recommendation error of the algorithms.

[0141] (2) When faced with datasets with different numbers of new users, both the TagMF algorithm and the UTagUser-TagMF algorithm have smaller recommendation errors than the MF algorithm, indicating that both algorithms can effectively solve the cold start problem and have better performance when facing the cold start phenomenon. Among them, the UTagUser-TagMF algorithm has a lower recommendation error than the TagMF algorithm when the number of new users is larger, indicating that the UTagUser-TagMF algorithm has a better ability to solve the cold start problem than the TagMF algorithm and has better recommendation performance when the number of new users increases.

[0142] In summary, the cold start problem affects algorithms, increasing recommendation error. Both TagMF and UTagUser-MF algorithms effectively address the cold start problem and exhibit good recommendation performance. UTagUser-MF, in particular, demonstrates superior cold start capability and is the preferred choice when dealing with a large number of new users.

[0143] Comparison with other algorithms:

[0144] The following text Figure 5 These are experimental data for three matrix factorization-based collaborative filtering algorithms at different iteration numbers. The iteration numbers (step) are 10, 20, 30, 40, and 50. The learning step size of the TagMF algorithm is also shown. The regularization coefficient is 0.002. The learning step size is 0.02. The regularization coefficient is 0.002. It is 0.02, combined with the weighting coefficient. It is 0.1.

[0145] observe Figure 5 It can be observed that the proposed TagMF algorithm and UTagUser-TagMF algorithm have lower recommendation errors than the MF-KNN algorithm and NMF algorithm when the number of iterations is 10, 20, 30, 40 and 50.

[0146] In summary, through the analysis of Figure 5 Based on observation and analysis, the following conclusions can be drawn from this experiment: The proposed TagMF algorithm and UTagUser-TagMF algorithm have lower recommendation errors and better recommendation performance. This indicates that the tag information model helps the matrix factorization model to increase hidden feature information, and incorporating the tag information model into matrix factorization can effectively reduce recommendation errors.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A recommendation algorithm based on a label neighborhood model and a label matrix factorization model, characterized in that: The UTagUser-CF algorithm and the TagMF algorithm are constructed, and a UTagUser-TagMF algorithm is constructed according to the constructed UTagUser-CF algorithm and the TagMF algorithm; First, the user's predicted score calculation formula of the UTagUser-CF algorithm is as formula (1); （1） represents the comprehensive similarity between users; Then, the user's predicted score calculation formula of the TagMF algorithm is as formula (2); （2） The UTa gUser-TagMF algorithm adopts a combined weight coefficient The two algorithms are combined, and the calculation formula of the prediction score is as shown in formula (3). （3） In formula (3), a predicted score representing the UTagUser-CF algorithm, a predicted score representing the TagMF algorithm; The TagMF algorithm is constructed as follows: In the tag information model, there are user tag model and project tag model. Assuming that the total number of project table tag types is T, the user tag model matrix is , and the project tag model matrix is . The two tag model matrices are combined and added to the original two initialization matrices, so that the two hidden feature matrices contain more context information, and the sizes of the two initialization matrices must be and , and the calculation formulas are as shown in equations (2-1) and (2-2), （2-1） （2-2） After the calculation of formula (2-1) and formula (2-2), two implicit feature matrices with label context information are obtained, and the two matrices are standardized by using the maximum and minimum standardization method, and the calculation formula is as formula (2-3): （2-3） In formula (2-3), a minimum value of the matrix, a maximum value of the matrix, each value in the matrix; After the first stage of initialization, the second stage is consistent with the commonly used MF algorithm, using the random gradient descent method to optimize the loss function, and constantly iterating the training, the optimal solution matrix can be obtained and ; Therefore, the TagMF algorithm is as formula (2) shown: （2）。 2.The recommendation algorithm based on a label neighborhood model and a label matrix factorization model according to claim 1, characterized in that: The UTagUser-CF algorithm is constructed as follows: The calculation steps of the label model are as follows: S1: a model is established for the target label of the project, a project label model matrix is constructed, it is assumed that the number of types of project labels is , a dimensional vector is established for each project, the label vector of the project is , it is assumed that the project is composed of label type 1 and label type 4, the 1st dimension and the 4th dimension in the project label vector are 1, and the rest are 0, then the value of the project label vector is as formula (1-1), （1-1） A tag vector is constructed for each item, and an item tag matrix ITag is generated, which has a size of where m is the number of items, is the number of tag types, and the matrix ITag has the form of equation (1-2), （1-2） S2: establishing a user tag model according to the project tag model and the scoring matrix, generating a dimensional vector for each user, and the tag vector of the user is , the dimension of the tag vector of the user is calculated by using the following formula, and the score value of the user for each type of tag is represented, and the calculation formula is as formula (1-3), ​ （1-3） The ITag matrix is an m-row T-column item label matrix, where m is the number of items, T is the number of label types, and 1 in the matrix indicates that the item is labeled with the label, In formula (1-3), denotes the label vector of user u, denotes the label model of item i, representing the item set scored by user u, a label vector is constructed for each user, and a user label matrix UTagis generated, the size of which is where n is the number of users, is the number of label types, and the matrix UTagis in the form of formula (1-4), （1-4） The UTag matrix is an n-row T-column matrix, where n is the number of users, T is the number of label types, and the matrix is generated by all user label vectors of formula (1-4); The label information model is integrated into the Neighborhood-CF algorithm to construct the UTagUser-CF algorithm; Combined with the user label model matrix UTag, each user has a label vector of his own, and for the label vector matrix of the user, the widely used Pearson coefficient can be used to calculate the similarity of the label vectors between users, that is, to evaluate the similarity of the users' preferences for label types. The user label vector similarity calculation formula is as formula (1-5), （1-5） In formula (1-5), representing a user and a user Similarity in the user tag model, representing the tag vector of user u, representing the average value of the tag vector of user u, representing the value of the i-th dimension of the tag vector of user u; After obtaining the label similarity of the user, an activation function is used to process the data of the user label similarity matrix, and the similarity less than the set threshold L in the user label similarity matrix is removed. The activation function is as formula (1-6), （1-6） representing a user and a user similarity in a user tag model; UTag matrix is used as the basis, and the similarity of the label model between users is calculated by formula (1-5), and then the data preprocessing of the label similarity matrix is performed by formula (1-6), and the obtained label similarity matrix is used as the similarity matrix. Such an algorithm is called a user label model-based collaborative filtering UTag-CF algorithm, which uses label information as data and can be used as a content-based recommendation algorithm, There are two different levels of user similarity, one based on the rating matrix and the other based on the user label model matrix. The two different user similarities are combined, and the combination formula of the two similarity matrices is as formula (1-7): （1-7） In formula (1-7), represents the comprehensive similarity between users, represents the user rating similarity, represents the user tag model similarity, represents the combination weight coefficient between the two, and the influence of the combination weight coefficient on the algorithm will be considered in subsequent experimental design. The comprehensive similarity matrix is used as the result of the first stage of the collaborative filtering algorithm, the second stage uses the number of neighbor sets or the comprehensive similarity threshold for neighbor selection, and the third stage uses formula (1) for the calculation formula of the predicted score. （1）。 3.The recommendation algorithm based on label neighborhood model and label matrix factorization model according to claim 1, characterized in that: After the UTagUser-TagMF algorithm is constructed, verification is needed, and the verification content includes the influence of different learning steps on the algorithm, the influence of different regularization coefficients on the algorithm, the influence of different combination weight coefficients on the algorithm, the influence of different data set sizes on the algorithm, the influence of the cold start problem on the algorithm, and the comparison with other algorithms.

Images