Neural collaborative filtering recommendation model considering multi-dimensional user preferences

Abstract

The invention provides a neural collaborative filtering recommendation model considering multi-dimensional user preferences, which comprises the following steps: acquiring and cleaning data information from a database to obtain an original user item score matrix; based on a Picture fuzzy set theory and a designed filling strategy, a user item scoring matrix is divided into four matrixes including a membership degree P, a non-membership degree N, a hesitation degree H and a rejection degree R so as to correspond to four interest preference states including user liking, disliking, hesitation and rejection; fitting the four preference matrixes by using a de-noising auto-encoder model, learning interests and preferences of the user in four dimensions, and constructing a user interest and preference model; performing normalization processing on the Picture fuzzy number (interests and preferences in four dimensions) of the user on the item obtained in the previous step, and obtaining an interests and preferences value of the user on the item through a scoring function; learning user behavior preferences by using a residual neural network fused with an attention mechanism, and generating a user behavior preference value by taking historical score information of a user on an article as input; and fusing the user interest preference and the user behavior preference in a weighting manner to form multi-dimensional user preference predicted values, and sorting the multi-dimensional user preference predicted values to generate a recommended item list.

Description

Technical Field

[0001] The present invention relates to the field of recommendation technology, and in particular to a neural collaborative filtering recommendation model that considers multi-dimensional user preferences. Background Art

[0002] The recommendation system can fully analyze the user's historical behavior information and build a corresponding recommendation model to predict user preferences and push the content that is most likely to interest them. This process effectively alleviates the problem of information overload and improves the user's experience of using the product. The collaborative filtering method based on deep neural networks has become one of the most popular nonlinear recommendation algorithms due to its powerful nonlinear transformation and learning capabilities, and can fully explore the low-order and high-order features contained in the data. It has good application potential.

[0003] However, since the data in the recommendation field is very sparse and the data of each user varies greatly, it is easy to have gradient anomalies when using deep learning models to capture the high-order features of the data. At the same time, when fusing the high-order and low-order features of the data, current research rarely considers the personalized preferences of users, and most studies ignore the impact of explicit information (such as ratings, comments, etc.) on the degree of user preference, making it impossible for the model to take into account the consistency and reliability of user purchase intention and interest preferences. Therefore, designing a neural collaborative filtering recommendation model that considers multi-dimensional user preferences, so that it can effectively fuse various features in the model and take into account the consistency and reliability of user purchase intention and interest preferences reflected in the data, is of great significance to improving the performance of the recommendation algorithm. Summary of the invention

[0004] In view of the above problems, the present invention proposes a neural collaborative filtering recommendation model that considers multi-dimensional user preferences. From the perspective of integrating user interest preferences and behavioral preferences, the concept of Picture fuzzy sets is introduced and combined with a denoising autoencoder to model user interest preferences. At the same time, a residual neural network based on a fusion attention mechanism is used to model user behavior preferences. Finally, interest preferences and behavioral preferences are integrated through weighted summation. Specifically, it includes the following:

[0005] The data collection and preparation module is used to obtain and clean the data collected in the database to obtain the required original scoring matrix.

[0006] The interest preference matrix division module is used to divide the user's preference degree. It introduces Picture fuzzy sets and divides the user's original rating matrix of the project into four matrices: membership P, non-membership N, hesitation H and rejection R, corresponding to the user's four preference states of like, dislike, hesitation and rejection.

[0007] The interest preference modeling module is used to model the user's preference status for items. It uses a denoising autoencoder to learn the feature combination of users and items in various preference states to obtain the user's predicted scores for items in four states.

[0008] The interest preference value calculation module is used to normalize the obtained user's image fuzzy number for the project (the predicted score under four states) and obtain the user's interest preference value for the project through a scoring function.

[0009] The behavior preference modeling module is used to learn the user's behavior preferences. It takes the user's historical behavior information as input and learns the user's behavior preferences through a residual neural network that integrates the attention mechanism.

[0010] The multi-dimensional user preference fusion module integrates the output values ​​of user interest preferences and user behavior preferences to obtain the user's final prediction value for the item.

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

[0012] The interest preference matrix partitioning module of the present invention is used to divide the user-project rating matrix into a membership matrix M, a non-membership degree N, a hesitation degree H, and a rejection matrix R. The rating of user u for project i is expressed as y u,i , the membership degree, non-membership degree, hesitation degree and rejection degree are m respectively u,i 、n u,i 、h u,i and r u,i Taking the scoring range of 1-5 points as an example, the setting rules of each interest preference matrix are as follows:

[0013] When u,i >3, the membership degree m u,i =1, accordingly, n u,i 、h u,i and r u,i 0, indicating that user u likes item i;

[0014] When u,i <3, non-membership degree n u,i =1, other values ​​are 0, indicating that user u does not like item i;

[0015] When u,i =3, the hesitation degree h u,i =1, and other values ​​are 0, indicating that user u is hesitant about item i.

[0016] For all unrated records, the following strategy is adopted to fill them. If the historical average score of item i is greater than 4 points, the corresponding position in the membership matrix is ​​filled with 1; for the 4-3 point interval, the corresponding position in the hesitation matrix is ​​filled with 1; for the 3-2 point interval, the corresponding position in the non-membership matrix is ​​filled with 1; for less than 2 points, the corresponding position in the rejection matrix is ​​filled with 1.

[0017] The interest preference modeling module of the present invention uses a denoising autoencoder model to parallel train the four interest preference matrices divided above to predict the interest preferences of the same blank area in each preference matrix. Taking a single interest preference modeling as an example, the denoising autoencoder model is as follows:

[0018] Record R of user u's interest preference in a certain state u As input, a single-layer fully connected neural network is used to obtain low-dimensional features, and the features of each dimension are fused in a linear weighted manner to generate a low-dimensional feature vector. The calculation formula is:

[0019]

[0020] Where h represents the low-dimensional feature vector; R u A vector representation abstracted from user history records; and b e They represent the weight parameters and bias parameters in the encoder respectively; σ represents the sigmoid function.

[0021] Subsequently, the decoder analyzes the extracted low-dimensional feature vectors, reconstructs the user's interest preferences, and completes the prediction of the user's preferences. The decoder is also composed of a single-layer fully connected neural network, and the calculation method is:

[0022]

[0023] Here represents the interest preference in a certain state after reconstruction, that is, the model prediction result; h represents the low-dimensional feature vector generated by the encoder; and b d They represent the weight parameters and bias parameters in the decoder respectively; σ represents the Sigmoid function.

[0024] Through the above denoising autoencoder, the membership matrix P, non-membership matrix N, hesitation matrix H and rejection matrix R are learned respectively to construct the feature mapping relationship of users in four dimensions, and finally the user's membership to the project is obtained. Non-membership Hesitation and rejection

[0025] The interest preference value calculation module of the present invention normalizes the predicted scores (Picture fuzzy numbers) of the user for the four states of the project obtained by the previous module through the Softmax function, and obtains the user's interest preference value for the project through the scoring function. The specific steps are as follows:

[0026] 1) Normalization processing, the specific calculation formula is as follows:

[0027]

[0028] 2) Use the scoring function to calculate the user's interest preference value for the project. The specific formula is as follows:

[0029]

[0030] The behavior preference modeling module described in the present invention learns the user's behavior preference by integrating the residual neural network with the attention mechanism. The specific steps are as follows:

[0031] 1) Establish a preprocessing layer to reconstruct the feature vector and improve the diversity of feature combinations. The specific calculation formula is as follows:

[0032]

[0033] Among them, V u,i represents the concatenation of user and item embedding vectors; X represents the output vector of MLP; Represents the output vector of the preprocessing layer, and is also the input vector of the next layer of the network; W p0 and W p1 represents the weight matrix of the preprocessing layer; b p0 and b p1 Represents the bias vector of the preprocessing layer.

[0034] 2) Residual channel design: By establishing a residual channel, the high-order feature vectors in the neural network can be fused with the lower-order feature vectors. At the same time, during the back-propagation process, the neural network updates the weight parameters through a shorter residual channel, so that the weights close to the input layer are effectively learned and the utilization of information close to the input layer is strengthened. The mathematical expression of the output of the nth (≥2) hidden layer in the network is:

[0035]

[0036] Among them, W n , b n , f n and h n is the weight matrix, bias vector, activation function and hidden layer output of the fully connected neural network, W i,n , b i,n and hi,n is the weight matrix, bias vector and hidden layer output of the residual channel of the i-th layer and the n-th layer.

[0037] 3) Attention mechanism layer design, considering that different feature combinations have different influences on the recommendation task, the attention mechanism model is introduced at the combination of the residual inputs of each layer. The attention layer calculates the attention score of each hidden layer input in a dot product manner. Its mathematical expression is shown in the formula:

[0038]

[0039] Among them, s n and i,n Respectively and h i,n The attention score of q represents the query vector related to the task in the attention mechanism.

[0040] According to the attention score, the attention mechanism layer calculates the attention distribution of the hidden layer through the Softmax function as shown in the formula):

[0041]

[0042] The inputs of each hidden layer are weighted and fused, and the final state of the hidden layer is obtained through the activation function. The mathematical expression is:

[0043]

[0044] 4) Calculation of behavior preference prediction value: The behavior preference prediction value of user u for item i is obtained through the residual neural network integrated with the attention mechanism mentioned above. The specific formula is as follows:

[0045]

[0046] Among them, l is the number of layers of the neural network.

[0047] The multi-dimensional user preference fusion module described in the present invention is used to combine user interest preferences and behavior preferences and perform training.

[0048] 1) Combine the output values ​​of user interest preference and user behavior preference by weighted summation to obtain the user's final prediction value for the item. The specific calculation formula is:

[0049]

[0050] Among them, ω1 and ω2 are weight parameters to determine the proportion of the prediction values ​​of the two modules in the final task.

[0051] 2) The specific formula of the model loss function is as follows:

[0052]

[0053] Among them, y u,i and Respectively represent the true value in the implicit feedback matrix and the predicted value after weighted fusion; W pre Represents all weight parameters in the behavior preference extraction module; λ pre represents the regularization coefficient.

[0054] The recommendation result acquisition module of the present invention is used to obtain the recommendation list of the target user u. The comprehensive prediction values ​​of the user's unrated items are sorted, and the top k items are selected to form a personalized recommendation list of the target user u.

[0055] The present invention proposes a neural collaborative filtering recommendation model that considers multi-dimensional user preferences. From the perspective of integrating user interest preferences and behavioral preferences, the concept of Picture fuzzy sets is introduced to divide user interest preferences, and a denoising autoencoder is used to model user interest preferences. At the same time, the residual neural network based on the fusion attention mechanism takes into account the memory and generalization capabilities of the model, ensures the effective transmission of user preference information in the network, and better models user behavior preferences. Finally, the interest preference and behavior preference are integrated through weighted summation to improve the recommendation quality of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of the model process of the present invention;

[0057] Figure 2 It is a model framework diagram of the present invention;

[0058] Figure 3 The Precision@10, Recall@10, F1@10 and NDCG@10 of the present invention are compared with other methods on three data sets. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention more clear, a neural collaborative filtering recommendation model considering multi-dimensional user preferences proposed by the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation method described herein is only used to explain the present invention and is not used to limit the present invention. Changes, modifications, additions, modifications or substitutions made by ordinary technicians in this technical field within the scope of the present invention should be included in the scope of the claims of the present invention.

[0060] Figure 1 This is a process diagram of a neural collaborative filtering recommendation model that considers multi-dimensional user preferences in the present invention. Figure 1It can be seen that the neural collaborative filtering recommendation model considering multi-dimensional user preferences proposed in the present invention includes acquiring and cleaning data information from a database to obtain an original user-item rating matrix; based on the Picture fuzzy set theory and the designed filling strategy, the rating matrix is ​​divided into four matrices representing the degree of user interest preference; the four preference matrices are fitted using a denoising autoencoder model to learn the user's interest preferences in four dimensions and construct a user interest preference model; the Picture fuzzy number (interest preference in four dimensions) of the user's item obtained in the previous step is normalized, and the user's interest preference value for the item is obtained through a scoring function; the user's behavior preference is learned using a residual neural network fused with an attention mechanism, and the user's historical rating information for the item is used as input to generate a user behavior preference matrix; the user interest preference and the user behavior preference are fused in a weighted manner to form a multi-dimensional user preference prediction value and sorted to generate a recommended item list.

[0061] Figure 2 This is a framework diagram of a neural collaborative filtering recommendation model that considers multi-dimensional user preferences in the present invention. Figure 2 It can be seen that the neural collaborative filtering recommendation model proposed in the present invention that takes into account multi-dimensional user preferences includes a data collection and preparation module, an interest preference matrix division module, an interest preference modeling module, an interest preference value calculation module, a behavior preference modeling module, a multi-dimensional user preference fusion module, and a recommendation result acquisition module. The data collection and preparation module is used to extract data from the database and perform operations such as cleaning and preprocessing on the data to obtain the required original user-item rating matrix; the interest preference matrix division module divides the user-item rating matrix into four matrices of membership P, non-membership N, hesitation H and rejection R according to the designed filling rules, corresponding to the four interest preference states of users: like, dislike, hesitation and rejection; the interest preference modeling module uses a denoising autoencoder, a neural network model with strong anti-interference ability and high efficiency, to learn the user's interest preferences in four dimensions; the interest preference value calculation module obtains the user's interest preference value for the project through a scoring function; the behavior preference modeling module takes the user's historical behavior information as input and learns the user's behavior preference through a residual neural network with an integrated attention mechanism; the multi-dimensional user preference fusion module weightedly sums the user's interest preference value and behavior preference to obtain the user's predicted value for the item; the recommendation result acquisition module is used to obtain the recommendation list of the target user u. The comprehensive predicted values ​​of the user's unrated items are sorted, and the top k items are selected to form a personalized recommendation list for the target user u.

[0062] Further, the following situation is used as an example to illustrate:

[0063] There are M users U={u1,u2,...,u M} and N items I={i1,i2,...,i N}, the rating domain is {1,2,3,4,5}, and the user-item rating matrix is ​​[R] M×N indicates; y ui represents the user's rating value on item i, U(i,j) represents the set of users who have rated both item i and item j, i.e., the common rating item; U i represents the set of all users who have evaluated item i.

[0064] 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 collaborative filtering recommendation model based on image fuzzy sets are as follows:

[0065] S1: The data collection and preparation module obtains the required data information from the database, including user ID, project ID and rating information.

[0066] S2: The interest preference matrix division module divides the user's original rating matrix of the project into four matrices: membership P, non-membership N, hesitation H and rejection R according to the Picture fuzzy set theory, which correspond to the user's four preference states of like, dislike, hesitation and rejection.

[0067] Taking the MovieLens dataset as an example, its rating domain is {1, 2, 3, 4, 5}, then the partitioning rule of the interest preference matrix is ​​as follows: the rating of user u for item i is represented by y u,i , the membership degree, non-membership degree, hesitation degree and rejection degree are m respectively u,i 、n u,i 、h u,i and r u,i For the scores recorded in the score matrix, when y u,i >3, the membership degree m u,i =1, accordingly, n u,i 、h u,i and r u,i are all 0, indicating that user u likes item i; when y u,i <3, non-membership degree n u,i =1, other values ​​are 0, indicating that user u does not like item i; when y u,i =3, the hesitation degree h u,i=1, and other values ​​are 0, indicating that user u is hesitant about item i. For ratings that are not recorded in the rating matrix, if the historical average score of item i is greater than 4 points, the corresponding position in the membership matrix is ​​filled with 1; for the 4-3 point range, the corresponding position in the hesitation matrix is ​​filled with 1; for the 3-2 point range, the corresponding position in the non-membership matrix is ​​filled with 1; for less than 2 points, the corresponding position in the rejection matrix is ​​filled with 1.

[0068] S3: The interest preference modeling module uses the denoising autoencoder to parallel train the four interest preference matrices obtained to predict the preference of the items to be tested in the preference matrix, namely the Picture fuzzy number.

[0069] Record R of user u's interest preference in a certain state u As input, a single-layer fully connected neural network is used to obtain low-dimensional features, and the features of each dimension are fused in a linear weighted manner to generate a low-dimensional feature vector. The calculation formula is:

[0070]

[0071] Where h represents the low-dimensional feature vector; R u A vector representation abstracted from user history records; and b e They represent the weight parameters and bias parameters in the encoder respectively; σ represents the sigmoid function.

[0072] Subsequently, the decoder analyzes the extracted low-dimensional feature vectors, reconstructs the user's interest preferences, and completes the prediction of the user's preferences. The decoder is also composed of a single-layer fully connected neural network, and the calculation method is:

[0073]

[0074] Here represents the interest preference in a certain state after reconstruction, that is, the model prediction result; h represents the low-dimensional feature vector generated by the encoder; and b d They represent the weight parameters and bias parameters in the decoder respectively; σ represents the Sigmoid function.

[0075] Through the above denoising autoencoder, the membership matrix P, non-membership matrix N, hesitation matrix H and rejection matrix R are learned respectively to construct the feature mapping relationship of users in four dimensions, and finally the user's membership to the project is obtained. Non-membership Hesitation and rejection

[0076] S4: The interest preference value calculation module normalizes the user's interest preference prediction scores (Picture fuzzy numbers) under the four states of the project obtained by the previous module through the Softmax function, and obtains the user's interest preference value for the project through the scoring function. The specific steps are as follows:

[0077] 1) Normalization processing, the specific calculation formula is as follows:

[0078]

[0079] 2) Use the scoring function to calculate the user's interest preference value for the project. The specific formula is as follows:

[0080]

[0081] S5: Behavior preference building module, which learns the user's behavior preference by integrating the residual neural network with the attention mechanism. The specific steps are as follows:

[0082] 1) Establish a preprocessing layer to reconstruct the feature vector and improve the diversity of feature combinations. The specific calculation formula is as follows:

[0083]

[0084] Among them, V u,i represents the concatenation of user and item embedding vectors; X represents the output vector of MLP; Represents the output vector of the preprocessing layer, and is also the input vector of the next layer of the network; W p0 and W p1 represents the weight matrix of the preprocessing layer; b p0 and b p1 Represents the bias vector of the preprocessing layer.

[0085] 2) Residual channel design: By establishing a residual channel, the high-order feature vectors in the neural network can be fused with the lower-order feature vectors. At the same time, during the back-propagation process, the neural network updates the weight parameters through a shorter residual channel, so that the weights close to the input layer are effectively learned and the utilization of information close to the input layer is strengthened. The mathematical expression of the output of the nth (≥2) hidden layer in the network is:

[0086]

[0087] Among them, W n , b n , f n and h n is the weight matrix, bias vector, activation function and hidden layer output of the fully connected neural network, W i,n , b i,n and hi,n is the weight matrix, bias vector and hidden layer output of the residual channel of the i-th layer and the n-th layer.

[0088] 3) Attention mechanism layer design, considering that different feature combinations have different influences on the recommendation task, the attention mechanism model is introduced at the combination of the residual inputs of each layer. The attention layer calculates the attention score of each hidden layer input in a dot product manner. Its mathematical expression is shown in the formula:

[0089]

[0090] Among them, s n and i,n Respectively and h i,n The attention score of q represents the query vector related to the task in the attention mechanism.

[0091] According to the attention score, the attention mechanism layer calculates the attention distribution of the hidden layer through the Softmax function as shown in the formula):

[0092]

[0093] The inputs of each hidden layer are weighted and fused, and the final state of the hidden layer is obtained through the activation function. The mathematical expression is:

[0094]

[0095] 4) Calculation of behavior preference prediction value: The behavior preference prediction value of user u for item i is obtained through the residual neural network integrated with the attention mechanism mentioned above. The specific formula is as follows:

[0096]

[0097] Among them, l is the number of layers of the neural network.

[0098] S6: Multi-dimensional user preference fusion module, used to combine user interest preferences and behavior preferences and perform training.

[0099] The output values ​​of user interest preference and user behavior preference are combined by weighted summation to obtain the user's final prediction value for the item. The specific calculation formula is:

[0100]

[0101] Among them, ω1 and ω2 are weight parameters to determine the proportion of the prediction values ​​of the two modules in the final task.

[0102] The specific formula of the model loss function is as follows:

[0103]

[0104] Among them, y u,i and Respectively represent the true value in the implicit feedback matrix and the predicted value after weighted fusion; W pre Represents all weight parameters in the behavior preference extraction module; λ pre represents the regularization coefficient.

[0105] S7: Recommendation result acquisition module, used to obtain the recommendation list of target user u. Sort the comprehensive prediction values ​​of the user's unrated items, and select the top k items to form a personalized recommendation list for target user u.

[0106] Figure 3 The proposed method is measured in four aspects on three datasets and compared with other methods. The four metrics are: Precision, Recall, F1 value, and Normalized Discounted Cumulative Gain (NDCG). Their calculation methods and measurement contents are as follows:

[0107] Precision represents the accuracy of the recommendation list, that is, the list of items recommended to the user. p (u i ) is the ratio of the items that users actually purchased to the recommended list. The calculation method is as follows:

[0108]

[0109] Recall represents the recall rate of the recommendation list, that is, the difference between the set of items in the list of items recommended to the user that the user actually purchased and the set of items that the user actually purchased. a (u i ) is calculated as follows:

[0110]

[0111] The F1 value is used as an evaluation indicator for recommendation ranking. The calculation method is as follows:

[0112]

[0113] Normalized Discounted Cumulative Gain (NDCG): This shows the quality of the model’s item ranking recommendations by assigning higher scores to top-ranked clicks to account for the click’s position. The calculation formula is as follows:

[0114]

[0115] Among them, DCG@K represents the cumulative loss of recommending K items to user u, rel idx∈{0,1} represents the relevance of the recommendation result at position idx. If the item is in the user's actual purchase item set, then rel idx The value is 1, otherwise it is 0. IDCG u @K is pressing rel idx Ideal DCG calculated on the recommendation list in descending order u @K. K represents the length of the recommendation list.

Claims

1. A neural collaborative filtering recommendation model that considers multi-dimensional user preferences is used to integrate user interest preferences and behavior preferences to improve the quality of recommendations, including the following: The data collection and preparation module is used to obtain and clean the data collected in the database to obtain the required original scoring matrix; The interest preference matrix division module is used to divide the user's preference degree. It introduces the Picture fuzzy set and divides the user's original rating matrix of the project into four matrices: membership P, non-membership N, hesitation H and rejection R, corresponding to the four preference states of the user: like, dislike, hesitation and rejection; The interest preference modeling module is used to model the user's preference status for items. It uses a denoising autoencoder to learn the feature combination of users and items in various preference states to obtain the user's predicted scores for items in four states. The interest preference value calculation module is used to normalize the obtained fuzzy number of the user's image of the project (the predicted score in four states) and obtain the user's interest preference value for the project through a scoring function; The behavior preference modeling module is used to learn the user's behavior preferences. It takes the user's historical behavior information as input and learns the user's behavior preferences through a residual neural network that integrates the attention mechanism; Multi-dimensional user preference fusion module, which integrates the output values ​​of user interest preference and user behavior preference to obtain the final prediction value of the user for the item; The recommendation result acquisition module sorts the comprehensive prediction values ​​of users for unrated items and selects the top k items to generate a recommended item list for the user.

2. A neural collaborative filtering recommendation model considering multi-dimensional user preferences as claimed in claim 1, characterized in that: The interest preference matrix partitioning module is used to divide the user-item rating matrix into membership matrix M, non-membership degree N, hesitation degree H, and rejection matrix R. The rating of user u on item i is expressed as y u,i , the membership degree, non-membership degree, hesitation degree and rejection degree are m respectively u,i 、n u,i 、h u,i and r u,i Taking the scoring range of 1-5 points as an example, the setting rules of each interest preference matrix are as follows: When u,i >3, the membership degree m u,i =1, accordingly, n u,i 、h u,i and r u,i 0, indicating that user u likes item i; When u,i <3, non-membership degree n u,i =1, other values ​​are 0, indicating that user u does not like item i; When u,i =3, the hesitation degree h u,i =1, and other values ​​are 0, indicating that user u is hesitant about item i. For all unrated records, the following strategy is adopted to fill them. If the historical average score of item i is greater than 4 points, the corresponding position in the membership matrix is ​​filled with 1; for the 4-3 point interval, the corresponding position in the hesitation matrix is ​​filled with 1; for the 3-2 point interval, the corresponding position in the non-membership matrix is ​​filled with 1; for less than 2 points, the corresponding position in the rejection matrix is ​​filled with 1.

3. The neural collaborative filtering recommendation model considering multi-dimensional user preferences according to claim 1, characterized in that: The interest preference modeling module uses a denoising autoencoder model to parallel train the four interest preference matrices divided above to predict the interest preferences of the same blank area in each preference matrix. Taking a single interest preference modeling as an example, the denoising autoencoder model is as follows: Record R of user u's interest preference in a certain state u As input, a single-layer fully connected neural network is used to obtain low-dimensional features, and the features of each dimension are fused in a linear weighted manner to generate a low-dimensional feature vector. The calculation formula is: Where h represents the low-dimensional feature vector; R u A vector representation abstracted from user history records; and b e They represent the weight parameters and bias parameters in the encoder respectively; σ represents the sigmoid function. Subsequently, the decoder analyzes the extracted low-dimensional feature vectors, reconstructs the user's interest preferences, and completes the prediction of the user's preferences. The decoder is also composed of a single-layer fully connected neural network, and the calculation method is: Here represents the interest preference in a certain state after reconstruction, that is, the model prediction result; h represents the low-dimensional feature vector generated by the encoder; and b d They represent the weight parameters and bias parameters in the decoder respectively; σ represents the Sigmoid function. Through the above denoising autoencoder, the membership matrix P, non-membership matrix N, hesitation matrix H and rejection matrix R are learned respectively to construct the feature mapping relationship of users in four dimensions, and finally the user's membership to the project is obtained. Non-membership Hesitation and rejection 4. The neural collaborative filtering recommendation model considering multi-dimensional user preferences according to claim 1, characterized in that: The interest preference value calculation module normalizes the predicted scores (Picture fuzzy numbers) of the user for the four states of the project obtained by the previous module through the Softmax function, and obtains the user's interest preference value for the project through the scoring function. The specific steps are as follows: S1: Normalization processing, the specific calculation formula is as follows: S2: Use the scoring function to calculate the user's interest preference value for the project. The specific formula is as follows:

5. The neural collaborative filtering recommendation model considering multi-dimensional user preferences according to claim 1, characterized in that: The behavior preference modeling module learns the user's behavior preference by integrating the residual neural network with the attention mechanism. The specific steps are as follows: S1: Establish a preprocessing layer to reconstruct the feature vector and improve the diversity of feature combinations. The specific calculation formula is as follows: Among them, V u,i represents the concatenation of user and item embedding vectors; X represents the output vector of MLP; Represents the output vector of the preprocessing layer, and is also the input vector of the next layer of the network; W p0 and W p1 represents the weight matrix of the preprocessing layer; b p0 and b p1 Represents the bias vector of the preprocessing layer. S2: Residual channel design, by establishing a residual channel, the high-order feature vectors in the neural network can be fused with the lower-order feature vectors. At the same time, during the back propagation process, the neural network updates the weight parameters through a shorter residual channel, so that the weights close to the input layer are effectively learned, and the utilization of information close to the input layer is strengthened. The mathematical expression of the output of the nth (≥2) hidden layer in the network is: Among them, W n , b n , f n and h n is the weight matrix, bias vector, activation function and hidden layer output of the fully connected neural network, W i,n , b i,n and h i,n is the weight matrix, bias vector and hidden layer output of the residual channel of the i-th layer and the n-th layer. S3: Attention mechanism layer design, considering that different feature combinations have different degrees of influence on the recommendation task, the attention mechanism model is introduced at the combination of the residual input of each layer. The attention layer calculates the attention score of each hidden layer input in a dot product manner. Its mathematical expression is shown in the formula: Among them, s n and i,n Respectively and h i,n The attention score of q represents the query vector related to the task in the attention mechanism. According to the attention score, the attention mechanism layer calculates the attention distribution of the hidden layer through the Softmax function as shown in the formula): The inputs of each hidden layer are weighted and fused, and the final state of the hidden layer is obtained through the activation function. The mathematical expression is: S4: Calculation of behavior preference prediction value. The behavior preference prediction value of user u for item i is obtained through the residual neural network integrated with the attention mechanism. The specific formula is as follows: Among them, l is the number of layers of the neural network.

6. A neural collaborative filtering recommendation model that considers multi-dimensional user preferences as described in claim 1, a multi-dimensional user preference fusion module, which is used to combine user interest preferences and behavior preferences and perform training. S1: Combine the output values ​​of user interest preference and user behavior preference by weighted summation to obtain the user's final prediction value for the item. The specific calculation formula is: in, ω1 and ω2 are weight parameters that determine the weight of the prediction values ​​of the two modules in the final task. S2: The specific formula of the model loss function is as follows: Among them, y u,i and Respectively represent the true value in the implicit feedback matrix and the predicted value after weighted fusion; W pre Represents all weight parameters in the behavior preference extraction module; λ pre represents the regularization coefficient.

7. The neural collaborative filtering recommendation model considering multi-dimensional user preferences according to claim 1, characterized in that: The recommendation result acquisition module is used to obtain the recommendation list of the target user u. The comprehensive prediction values ​​of the user's unrated items are sorted, and the top k items are selected to form a personalized recommendation list of the target user u.

Images