Picture fuzzy set-based collaborative filtering recommendation model
By using a collaborative filtering recommendation model based on image fuzzy sets, the problems of fuzziness and diversity of user preferences are solved. By obtaining the fuzzy number of images and learning global user preferences through a multilayer perceptron, the accuracy and adaptability of the recommendation system are improved.
Patent Information
- Application Number
- CN202310884977.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-18
- Publication Date
- 2025-11-28
- Estimated Expiration
- 2043-07-18
AI Technical Summary
Existing collaborative filtering recommendation algorithms fail to effectively handle the ambiguity and diversity of user preference information, leading to a decline in recommendation quality.
A collaborative filtering recommendation model based on image fuzzy sets is adopted. Through modules such as data collection, preference matrix generation, image fuzzy number acquisition, user global preference matrix generation, image fuzzy number scoring, and comprehensive prediction value calculation, the model combines multilayer perceptron to learn users' global preferences, thereby improving the accuracy of the recommendation model.
This improves the adaptability and accuracy of the recommendation model, enabling it to better match the uncertainty and diversity of user preferences and thus enhance the quality of recommendations.
Smart Images
Figure CN117009675B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the recommendation technical field, and particularly relates to a collaborative filtering recommendation model based on picture fuzzy sets. BACKGROUND
[0002] In the era of big data, recommendation models, as an important decision support tool, can predict the possibility of users accepting recommended goods based on the historical behavior information (such as evaluation, click or browse goods, etc.) of users, thereby effectively alleviating the problem of information overload. Collaborative filtering has become one of the mainstream recommendation technologies due to its easy engineering implementation and good recommendation quality.
[0003] Most of the existing collaborative filtering recommendation algorithm researches learn user preference information from a quantitative perspective by uniformly converting the historical ratings of users. However, due to the uncertainty and complexity of the decision environment, the preferences of users for items often have fuzziness and diversity. In the recommendation system, users express their preference information for items through ratings, but users may have difficulty in accurately expressing the preference degree for items, that is, the user preference has fuzziness. At the same time, a user may like multiple types of items at the same time, that is, the user preference has diversity. However, many studies uniformly convert the historical behavior of users into user preference information and learn user preference from a quantitative perspective using rating information, ignoring the uncertainty and diversity of user preference, thereby reducing the recommendation quality of the model. Therefore, in order to more accurately learn the preference information of users, a collaborative filtering recommendation model based on picture fuzzy sets is designed, which can learn the fine-grained preference information and global preference information of users at the same time. SUMMARY
[0004] In view of the above problems, in order to take into account the fuzziness and diversity of user preference, the present application proposes a collaborative filtering recommendation model based on picture fuzzy sets from the perspective of user preference fuzzy probability, which includes the following:
[0005] The data acquisition and preparation module is used for obtaining and cleaning the collected data in the database to obtain the required original rating matrix;
[0006] The preference matrix generation module is used for dividing the user preference degree, and based on the picture fuzzy set theory, the original rating matrix R of the user and the item is converted into four different preference matrices containing only 0 and 1 according to the preference score, that is, the membership matrix M, the neutral matrix N, the non-membership matrix H and the rejection matrix W;
[0007] The picture fuzzy number acquisition module is used for acquiring the picture fuzzy number of the user for the item, and for the above four preference matrices, the BeMF model is used for fitting respectively, so as to predict the picture fuzzy number of the user for the item;
[0008] A picture blur number scoring module is configured to quantitatively compare the picture blur number of the user to the item, and score the obtained picture blur number by using a scoring function.
[0009] A user global preference mining module is configured to learn the global preference of the user, and take the historical behavior information of the user as input, and learn the global preference of the user by using a multilayer perception machine to supplement the preference information of the user.
[0010] A comprehensive prediction value calculation module is configured to integrate the output values of the user preference fine-grained mining module and the user global preference mining module to obtain the prediction value of the user to the item.
[0011] A recommendation result acquisition module is configured to sort the comprehensive prediction value of the user to the un-rated item, and select the top k items to generate a recommendation item list of the user.
[0012] The preference matrix generation module is configured to divide the user-item rating matrix R into corresponding several independent and only containing 0-1 binary sub-matrices, i.e., membership matrix M, neutral matrix N, non-membership matrix H and rejection matrix W. Taking the score interval of 1-5 as an example, the setting rules of the preference matrix are as follows:
[0013] If the score of the user i to the item j is greater than 3, it is considered that the user i likes the item j and M ij = 1.
[0014] If the score of the user i to the item j is equal to 3, it is considered that the preference degree of the user i to the item j is neutral and N ij = 1.
[0015] If the score of the user i to the item j is less than 3, it is considered that the user i does not like the item j and H ij = 1.
[0016] In all un-rated records, a certain number of user items are extracted to construct the rejection matrix W.
[0017] The four preference matrices are related to each other, if W ij = 1, N ij = 0, H ij = 0, M ij = 0. If a position in the preference matrix is empty, i.e., the position is a blank area to be predicted by the model.
[0018] The picture blur number acquisition module uses the BeMF model to train the obtained preference matrix in parallel to predict the picture blur number of the same blank area in each preference matrix. For the four preference matrices M, N, H and W, the user feature matrix and the item feature matrix pairs are learned, respectively, as (P M , Q M ), (PN Q N ),(P H Q H ),(P W Q W Specifically, the steps include the following:
[0019] S1: In the BeMF model training process, the sigmoid function is selected as the smooth increasing function, i.e., g(x) = 1 / (1+e^x). -x Therefore, the feature vectors of user i and item j are updated as follows:
[0020]
[0021]
[0022] Where γ and η represent the regularization coefficient and learning rate, respectively. The corresponding feature matrix is updated in the manner described above until the model converges.
[0023] S2: Image blur count prediction, using the inner product of eigenvectors to predict the image blur count α = (μ) for user i on item j. a (r),η a (r),υ a (r),ω a (r)), which represent the probabilities that user i has different levels of preference for item j. Considering that the inner product of the feature vectors may not necessarily sum to 1, the softmax function is used to normalize it, that is:
[0024]
[0025] The image blur count scoring module described in this invention is used to score the blur count of images. A scoring function s(α) is defined for the image blur set. For two image blur counts α and β, if s(α) > s(β), then α is superior to β. The image blur counts can be sorted according to this rule. The scoring function is as follows:
[0026] S(a)=k1*u a (r)+k2*η a (r)+k3*υ a (r)-k4*ω a (r)
[0027] Where k1, k2, k3, and k4 are weight coefficients. If their values are all 1, it means that different levels of preference are assigned the same weight.
[0028] The user global preference mining module learns the global preference of the user through a multi-layer perception machine to supplement the preference information of the user and improve the expression capability of the model. i , the historical rating information of the item j is V j , and the specific steps are as follows:
[0029] S1: learning user preference information and item audience information Taking the user preference information as an example, the specific calculation formula is as follows:
[0030]
[0031] ...
[0033]
[0034] Wherein H is the number of layers of the multi-layer perception machine, W H ,b H ,σ H respectively represent the weight matrix, the bias vector and the activation function of the Hth layer. The item audience information is learned in the same way;
[0035] S2: output the predicted value of the user i to the item j through the inner product of the user preference information and the item audience information , and the specific formula is as follows:
[0036]
[0037] Wherein σ is a sigmoid function;
[0038] S3: the loss function of the model, the loss function is constructed based on the cross-entropy loss, and L2 regularization is used to prevent the model from overfitting. The calculation formula is as follows:
[0039]
[0040] Wherein Θ is the weight parameter of the multi-layer perception machine, Y,Y - respectively represent the positive and negative sample sets of the model training, Y ij , respectively represent the true value and the predicted value, and λ is the regularization term coefficient.
[0041] The comprehensive predicted value calculation module integrates the score of the picture fuzzy number and the output value of the user global preference mining module to obtain the predicted value of the user to the item, and the specific calculation formula is as follows:
[0042]
[0043] Wherein, η1, η2 are weight coefficients of two modules respectively.
[0044] The recommendation result acquisition module is used for acquiring a recommendation list of a target user u.
[0045] The collaborative filtering recommendation model based on the picture fuzzy set proposed in the application qualitatively divides the user preference by means of the picture fuzzy set theory, then learns the fine-grained preference of the user through the BeMF model, so as to better match the uncertainty of the user preference, and improve the sorting rule of the picture fuzzy number to improve the adaptability of the picture fuzzy number in the recommendation system. BRIEF DESCRIPTION OF DRAWINGS
[0046] Figure 1 is a model process schematic diagram of the application;
[0047] Figure 2 is a model framework diagram of the application;
[0048] Figure 3 is the HR@10 value and NDCG@10 value of the application under three data sets and the comparison of other methods. DETAILED DESCRIPTION
[0049] In order to make the purposes, technical solutions and advantages of the application clearer, the application of a collaborative filtering recommendation model based on picture fuzzy set is described in detail below in combination with the drawings. It should be understood that the specific implementation methods described herein are only used to explain the application, and are not used to limit the application, and the changes, modifications, additions, or replacements made by those skilled in the art within the scope of the application should be covered in the scope of the claims of the application.
[0050] Figure 1 is a process schematic diagram of the collaborative filtering recommendation model based on picture fuzzy set of the application. From Figure 1It can be seen that the collaborative filtering recommendation model based on picture fuzzy set provided by the present application comprises the following steps: obtaining and cleaning data information from a database to obtain a required user-item original rating matrix; based on picture fuzzy set theory, converting the rating matrix into four matrices representing different preference degrees of users, namely membership matrix, neutral matrix, non-membership matrix and rejection matrix; using BeMF model to fit the four preference matrices to obtain latent feature vectors of users and items, and then determining picture fuzzy numbers of users to items through inner product of the feature vectors and normalizing the picture fuzzy numbers; scoring the picture fuzzy numbers by means of a scoring function; learning global preference of users by using a multi-layer perception, taking historical rating information of users to items as input, learning feature vector representation of users and items through the multi-layer perception, and outputting inner product of the feature vectors as global preference information of users to items; combining the score of the picture fuzzy numbers (fine-grained preference) and the output value of the multi-layer perception (global preference) to form a comprehensive prediction value and sort the comprehensive prediction value to generate a recommended item list.
[0051] Figure 2 is a framework diagram of the collaborative filtering recommendation model based on picture fuzzy set of the present application. From Figure 2 It can be seen that the collaborative filtering recommendation model based on picture fuzzy set provided by the present application comprises a data acquisition and preparation module, a preference matrix generation module, a picture fuzzy number acquisition module, a picture fuzzy number scoring module, a user global preference mining module, a comprehensive prediction value calculation module and a recommended result acquisition module. The data acquisition and preparation module is used to extract data from a database and perform cleaning and preprocessing operations on the data to obtain a required original rating matrix; the preference matrix generation module divides a user-item rating matrix R into corresponding four preference matrices containing only 0-1 binary values according to system settings; the picture fuzzy number acquisition module uses BeMF model to train the acquired preference matrices in parallel to predict the same blank area value in each preference matrix, i.e. picture fuzzy numbers; the picture fuzzy number scoring module is used to score and sort the picture fuzzy numbers; the user global preference mining module takes historical behavior information of users as input to learn global preference of users through a multi-layer perception to supplement preference information of users and improve expression ability of the model; the comprehensive prediction value calculation module integrates the score of the picture fuzzy numbers and the output value of the user global preference mining module to obtain a prediction value of users to items; and the recommended result acquisition module is used to obtain a recommended list of a target user u. The comprehensive prediction values of items not rated by the user are sorted, and the first k items are selected to form a personalized recommendation list of the target user u.
[0052] Further, the following cases are taken as examples for illustration:
[0053] The MovieLens dataset contains M users U = {u1, u2,..., uM} and N items I = {i1, i2,..., iN}. The user-item rating matrix R is a sparse matrix of size M x N, where Rui is the rating of user ui to item ij. The rating scale is from 1 to 5. M} and N items I = {i1, i2,..., i N} and the rating domain is {1, 2, 3, 4, 5}, the user-item rating matrix is denoted as [R] M×N where r ui denotes the rating value of user i on item j.
[0054] Firstly, the system data is obtained and cleaned by the data acquisition and preparation module to obtain each variable in the hypothesis. The specific implementation steps of the whole picture fuzzy set based collaborative filtering recommendation model are as follows:
[0055] S1: The data acquisition and preparation module obtains the required data information from the database, including user ID, item ID and rating information.
[0056] S2: The preference matrix generation module converts the user-item rating matrix R into four different preference matrices containing only 0 and 1 according to the preference score based on the picture fuzzy set theory, i.e. the membership matrix M, the neutral matrix N, the non-membership matrix H and the rejection matrix W.
[0057] Taking the MovieLens dataset as an example, the rating domain is {1, 2, 3, 4, 5}, and the setting rules of the preference matrix are as follows: if the rating of user i on item j is greater than 3, it is considered that user i likes item j and M ij = 1; if the rating of user i on item j is equal to 3, it is considered that the preference degree of user i on item j is neutral and N ij = 1; if the rating of user i on item j is less than 3, it is considered that user i does not like item j and H ij = 1; in all un-rated records, a certain number of user items are extracted to construct the rejection matrix W. The four preference matrices are related to each other, if W ij = 1, then N ij = 0, H ij = 0, M ij = 0. If a position in the preference matrix is empty, i.e. the position is a blank area to be predicted by the model.
[0058] S3: The picture fuzzy number acquisition module uses BeMF to train the four obtained preference matrices in parallel to predict the preference of the test item in the preference matrix, i.e. the picture fuzzy number.
[0059] 1) BeMF model training process, the feature vector updating method of user i and item j is:
[0060]
[0061]
[0062] where γ,η represent the regularization term coefficient and learning rate respectively.
[0063] 2) Picture blur number prediction, using the inner product of feature vectors to predict the picture blur number of user i to item j, α = (μ a (r),η a (r),υ a (r),ω a (r)) respectively represent the probability of user i to item j belonging to different preference levels. And use the softmax function to normalize it, that is:
[0064]
[0065] S4: Picture blur number scoring module, used to score the picture blur number. The scoring function is as follows:
[0066] S(a) = k1*u a (r) + k2*η a (r) + k3*υ a (r) - k4*ω a (r)
[0067] where k1, k2, k3, k4 are weight coefficients, giving different weights to different preference levels, and their values are the mean of preference scores, to reflect the difference between different preference levels.
[0068] S5: User global preference mining module, let the historical rating information vector of user i be V i , and the historical rating information of item j be V j , then the framework of multi-layer perception learning user global preference information representation is:
[0069]
[0070] ...
[0072]
[0073] where H is the number of layers of multi-layer perception, W H ,b H ,σ H represent the weight matrix, bias vector and activation function of the Hth layer respectively. Item audience information representation is learned in the same way. Finally, the inner product of user preference information and item audience information is used to output the predicted value of user i to item j, and the calculation formula is as follows:
[0074]
[0075] where σ is sigmoid function. Considering the problem of overfitting, a regularization term is added, and the loss function of the model is:
[0076]
[0077] where Θ is the weight parameter of the multi-layer perception, Y, Y - represent the positive and negative sample sets of model training, Y ij , represent the true value and predicted value respectively, and λ is the regularization term coefficient. The model uses the Adam optimizer to optimize the model parameters.
[0078] S6: The comprehensive predicted value calculation module integrates the output values of the picture blur number scoring and user global preference mining modules to obtain the predicted value of the user for the item, and the specific calculation formula is:
[0079]
[0080] where η1, η2 are the weight coefficients of the two modules respectively.
[0081] S7: The recommendation result acquisition module is used to obtain the recommendation list of the target user u. The comprehensive predicted values of the items that the user has not scored are sorted, and the top k items are selected to form the personalized recommendation list of the target user u.
[0082] Figure 3 The performance of the proposed method in three data and two aspects of indicators is measured, and compared with other methods. The two measurements are: hit rate (HR) and normalized discounted cumulative gain (NDCG). Their calculation methods and measurement contents are as follows:
[0083] Hit rate (HR): used to verify whether the test item exists in the recommended item list, which indicates the item recommendation ability of the model, and the calculation method is as follows:
[0084]
[0085] where I p (u)@K represents the top K items in the recommended item list of user u, I a (u) represents the items actually interacted by user u.
[0086] Normalized discounted cumulative gain (NDCG): by assigning higher scores to the top clicks, it shows the position of the clicks, which shows the quality of the item ranking recommendation of the model, and the calculation formula is as follows:
[0087]
[0088]
[0089] Where DCG@k represents the cumulative gain from recommending K items to user u, rel idx ∈{0,1} represents the relevance of the recommendation result for location idx. If the item is in the user's actual purchased item set, then rel idx The value is 1 otherwise. IDCG u @K is pressing rel. idx The ideal DCG calculated on the descending list of recommendations u @K. K represents the length of the recommendation list.
Claims
1. A picture fuzzy set based collaborative filtering recommendation model for simultaneously learning user's fine-grained preference information and global preference information to improve recommendation quality, comprising the following: a data collection and preparation module for obtaining and cleaning the collected data in the database to obtain the required original rating matrix; a preference matrix generation module for dividing the user preference degree, converting the original rating matrix R of the user and the item into four different preference matrices with values containing only 0 and 1, i.e., membership matrix M, neutral matrix N, non-membership matrix H and rejection matrix W, based on the picture fuzzy set theory; a picture fuzzy number acquisition module for acquiring the picture fuzzy number of the user for the item, using the BeMF model to fit the above four preference matrices respectively to predict the picture fuzzy number of the user for the item; a picture fuzzy number scoring module for quantitatively comparing the picture fuzzy number of the user for the item, and scoring the obtained picture fuzzy number using a scoring function; a user global preference mining module for learning the global preference of the user, taking the historical behavior information of the user as input, and learning the global preference of the user through a multi-layer perception machine to supplement the preference information of the user; a comprehensive prediction value calculation module for integrating the output values of the user preference fine-grained mining module and the user global preference mining module to obtain the prediction value of the user for the item; a recommendation result acquisition module for sorting the comprehensive prediction value of the user for the un-rated item and selecting the top k items to generate a recommended item list for the user.
2. The picture set based collaborative filtering recommendation model of claim 1, wherein, The preference matrix generation module is used for dividing the user preference degree, converting the original rating matrix R of the user and the item into four different preference matrices with values containing only 0 and 1, i.e., membership matrix M, neutral matrix N, non-membership matrix H and rejection matrix W, based on the picture fuzzy set theory, taking the rating interval of 1-5 points as an example, and the setting rules of the preference matrix are as follows: If the rating of user i to item j is greater than 3, it is considered that user i likes item j and M ij = 1; If the rating of user i to item j is equal to 3, it is considered that the preference degree of user i to item j is neutral and N ij = 1; If the rating of item j by user i is less than 3, then user i is considered to dislike item j and H ij = 1. extract a certain number of user items from all un-rated records to construct the rejection matrix W, The four preference matrices are interrelated, if W ij = 1, then N ij = 0, H ij = 0, M ij = 0, if a position in the preference matrix is empty, i.e. the position is a blank area to be predicted by the model.
3. The picture-set-based collaborative filtering recommendation model of claim 1, wherein, The picture blur number acquisition module utilizes the BeMF model to train the obtained preference matrix in parallel to predict the normalized picture blur number of the same blank area in each preference matrix, and learns the user feature matrix and item feature matrix pairs of four preference matrices M, N, H, and W, respectively, as (P M , Q M , )(P N , Q N , )(P H , Q H , )(P W , Q W ), and specifically includes the following steps: S1: BeMF model training process, select a smooth increasing function as sigmoid function, that is, g(x) = 1 / (1+e -x Therefore, the feature vector updating method of user i and item j is: where γ and η represent the regularization term coefficient and the learning rate respectively, the corresponding feature matrix is updated in the above manner until the model converges; S2: picture blur number prediction and normalization, using the inner product of feature vectors to predict the picture blur number of user i to item j α = (μ a (r),η a (r),υ a (r),ω a (r)) respectively represent the probability of user i to item j belonging to different preference levels, considering that the inner product of the feature vectors does not necessarily satisfy the sum of 1, the softmax function is used for normalization operation, that is: .
4. The picture-set-based collaborative filtering recommendation model of claim 1, wherein, The picture fuzzy number scoring module is used for scoring the picture fuzzy number, the picture fuzzy set defines a scoring function s(α), for two picture fuzzy numbers α and β, if s(α) > s(β), it means that α is better than β, according to this rule, the picture fuzzy numbers can be sorted, and the scoring function is defined as follows: S(a) = k1 * u a (r) + k2 * η a (r) + k3 * v a (r) - k4 * ω a (r).
5. The picture-set-based collaborative filtering recommendation model as claimed in claim 1, wherein, The user global preference mining module learns the global preference of the user by using a multi-layer perception machine to supplement the preference information of the user and improve the expression capability of the model, the historical rating information vector of the user i is V i , and the historical rating information of the item j is V j . The specific steps are as follows: S1: learn user preference information with item audience information Taking the user preference information as an example, the specific calculation formula is as follows: ... where H is the number of layers of the multilayer perceptron, W H ,b H ,σ H represent the weight matrix, bias vector, and activation function of the Hth layer, respectively, and the item-audience information is represented in the same way; S2: by user preference information with item audience information the inner product of the user i's preference information and the item j's audience information to output the predicted value of the user i to the item j, the specific formula is as follows: where σ is a sigmoid function; S3: the loss function of the model, the loss function is constructed based on the cross-entropy loss, and L2 regularization is used to prevent model overfitting, and the calculation formula is as follows: where Θ is the weight parameter of the multi-layer perceptron, Y, Y - represent the positive and negative sample sets for model training, respectively, Y ij , represent the true value and predicted value, respectively, and λ is the regularization term coefficient. 6.The picture fuzzy set based collaborative filtering recommendation model of claim 1, wherein the comprehensive prediction value calculation module integrates the scoring of the picture fuzzy number and the output value of the user global preference mining module to obtain the prediction value of the user for the item, and the specific calculation formula is as follows: where η1 and η2 are the weight coefficients of the two modules.
7. The picture-set-based collaborative filtering recommendation model as claimed in claim 1, wherein, The recommendation result obtaining module is configured to obtain a recommendation list of a target user u, sort comprehensive prediction values of items that have not been scored by the user, and select the first k items to form a personalized recommendation list of the target user u.