A neural collaborative filtering recommendation method and system based on Chinese review mining

Abstract

The present invention relates to an aspect neural collaborative filtering recommendation method and system based on Chinese review mining, comprising: preprocessing a Chinese review data set; formulating aspect extraction rules for comments according to dependency syntax and part-of-speech rules to extract aspect terms; sending the extracted aspects into a latent Dirichlet distribution model for aspect clustering and converting them into potential factors; integrating user, item and aspect information contained in the Chinese review data set to construct a potential interaction feature tensor; sending the potential interaction feature tensor into an ANCF model for training and predicting scores; the present invention can perform unsupervised aspect extraction on Chinese reviews, and then combine aspect information into recommendations with low network complexity, pay attention to interaction information, reduce the impact of data sparsity, and complete accurate and efficient recommendations.

Description

Technical Field

[0001] The present invention belongs to the field of recommendation algorithms based on deep learning, and in particular relates to an aspect neural collaborative filtering recommendation method and system based on Chinese review mining. Background Art

[0002] With the explosive growth of online information, recommendation systems play a vital role in solving the problem of information overload. Collaborative filtering methods can make recommendations by using the ratings given by a large number of users. Data sparsity seriously affects the effectiveness and accuracy of recommendation systems, exacerbating the difficulties faced when introducing multidimensional information.

[0003] Considering that user text comments contain information about whether the user is interested in an item, recommendation methods have evolved from simply using user ratings to making recommendations by mining the information expressed in the comments. Furthermore, unlike predicting the total rating and user's overall preference based on document-level or sentence-level sentiment analysis, recommendations based on aspect information can predict detailed user preferences at the aspect level in a fine-grained manner. However, the introduction of aspect information in recommendation faces the problems of efficiency of aspect extraction and model applicability. For example, aspect extraction based on comment mining mostly relies on the construction of dictionaries or machine learning models. The former relies on time-consuming and labor-intensive dictionary construction and is not universal for Chinese comments, while the latter faces huge training costs. Recommendation systems that combine aspects often only use the method of cumulative model modules to introduce aspect information, which increases the complexity of the model and cannot well integrate multi-dimensional information and realize feature extraction, and it is difficult to solve the problem caused by data sparsity. Recommendation algorithms based on tensor decomposition are mostly based on context and lack a model that combines aspect features into tensor decomposition. Its disadvantage is that while increasing the data dimension, it also increases the sparsity of the data, making data fitting more difficult. In terms of neural recommendation algorithms that combine aspects, most existing models use the attention mechanism to introduce aspects into the recommendation layer. Although it alleviates the problem of data sparsity in the former to a certain extent, it increases the complexity of the model due to the accumulation of the model. At the same time, this method rarely pays attention to the contribution of both the potential features of the aspects and the actual interactions, and faces difficulties in combining the two.

[0004] At present, aspect-based recommendation systems face two problems: in terms of aspect extraction based on Chinese review mining, there is a lack of aspect extraction models that are generally applicable to Chinese reviews, and the existing models have poor adaptability and high cost, and are not suitable for recommendation systems that require large-scale data; in terms of aspect recommendation algorithms, there is a lack of models that integrate aspects into recommendations in a deep learning manner. The existing models are highly complex and cannot make good use of aspects to alleviate the cold start problem caused by data sparsity, and lack the utilization of the potential feature contributions of aspects. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an aspect-based neural collaborative filtering recommendation method based on Chinese review mining. Compared with the existing aspect-based collaborative filtering method, this method combines Chinese dependency syntactic analysis with feature word part-of-speech screening to extract aspects, and has universal applicability for accurate extraction of Chinese reviews; based on generalized tensor factorization (GTF) and tensor-multilayer perceptron (T-MLP), the potential features of users, items, and aspects are fully combined, overcoming the inherent data sparsity of traditional tensor factorization combined with context, which can well alleviate the cold start problem and improve the accuracy and interpretability of recommendations.

[0006] The present invention also aims to provide an aspect neural collaborative filtering recommendation system based on Chinese review mining.

[0007] Terminology explanation:

[0008] Latent Dirichlet allocation (LDA) is a topic model that can present the topic of each document in a document set in the form of probability distribution. It is an unsupervised learning algorithm that does not require manually annotated training sets during training. It only requires a document set and the number of specified topics. Another advantage of LDA is that for each topic, some words can be found to describe it.

[0009] The technical solution of the present invention is as follows:

[0010] An aspect-based neural collaborative filtering recommendation method based on Chinese review mining, including:

[0011] Preprocess and segment the Chinese review dataset;

[0012] Formulate aspect extraction rules for comments based on dependency syntax and part-of-speech rules to extract aspect terms;

[0013] After removing stop words from the extracted aspect terms, they are fed into the latent Dirichlet distribution model for aspect clustering and converted into latent factors to construct a Chinese review-aspect dataset;

[0014] Integrate the user, item, and aspect information contained in the Chinese review-aspect dataset to construct a potential interaction feature tensor;

[0015] Send the potential interaction feature tensor to the ANCF model for training to obtain the trained ANCF model;

[0016] The Chinese review-aspect dataset is fed into the trained ANCF model in the form of user id-item id-aspect id to obtain the recommendation results, where the aspect id is 1, representing the overall interaction between the user and the item.

[0017] Preferably, according to the present invention, the Chinese review data set is preprocessed, including: deleting user comments with poor sentence structure, fake reviews, malicious reviews, and fewer than a certain number of comments.

[0018] Preferably, according to the present invention, a comment aspect extraction rule is formulated according to dependency syntax and part-of-speech rules to perform comment mining, and the extracted aspects are saved as a comment aspect thesaurus; including:

[0019] Assume that there are N Chinese reviews, and perform the following steps (1) to (5) for each Chinese review:

[0020] (1) Select the nth Chinese comment and perform word segmentation on it;

[0021] (2) Perform dependency syntactic analysis to obtain all the subjects W pointed to by the subject-predicate relationship in the Chinese comments s n , and with the subject W s n W is a word in parallel relationship c n ;

[0022] (3) Get m dependency syntax extraction words:

[0023] (4) Filter the words in the Chinese comments by their parts of speech and extract all nouns and gerunds Names

[0024] (5) Get the final aspect word

[0025] Preferably, according to the present invention, after removing stop words, the extracted aspects are sent to the latent Dirichlet distribution model for aspect clustering and converted into latent factors, which means: the final aspect words obtained are sent to the latent Dirichlet distribution model for clustering, the number of clusters is set to L, the correlation index λ, and the aspect term group is obtained.

[0026] According to the preferred embodiment of the present invention, the user, item, and aspect information contained in the Chinese review dataset is integrated to construct a potential interaction feature tensor, which means: let M, N, and L represent the number of users, items, and aspects respectively, and the one-hot encoded user-item-aspect potential interaction tensor Definition formula (I):

[0027]

[0028] Among them, y ijk is an element of the potential interaction tensor Y, where a value of 1 indicates that there is an interaction between user i and aspect k of item j, and a value of 0 indicates that there is no interaction.

[0029] Preferably, according to the present invention, the ANCF model includes an input layer, an embedding layer, a generalized tensor decomposition layer, a tensor multilayer perceptron layer, a neural collaborative filtering layer, and an output layer;

[0030] In the input layer, M user IDs, N item IDs, and L aspect IDs are one-hot encoded and mapped to user, item, and aspect feature vectors respectively. And the three sets of feature vectors They are fed into the embedding layer respectively;

[0031] In the embedding layer, the embedding layer is a fully connected layer that constructs the three sets of feature vectors in the input layer. Mapped to a low-dimensional dense feature representation vector with a vector dimension of r, perform potential mapping encoding, and user potential interaction feature tensor Item potential interaction feature tensor Aspect latent interaction feature tensor

[0032] In the generalized tensor decomposition layer, the extended generalized tensor decomposition method is used to perform tensor decomposition on the potential interaction feature tensor;

[0033] In the tensor multilayer perceptron layer, a hidden layer is added to the connection vector inside the perceptron, and a multilayer perceptron that flattens the tensor is used to learn the interaction between the latent features of users, items, and aspects, and to understand the interaction between the latent feature vectors of users, items, and aspects with a nonlinear model;

[0034] In the neural tensor decomposition layer, the outputs of the tensor multi-layer perception layer and the generalized tensor decomposition layer are concatenated and fed back to the neural collaborative filtering layer to predict the score using the latent vector;

[0035] In the output layer, the likelihood function is used to constrain the prediction scores to [0,1]:

[0036] Further preferably, the activation function of the output layer is a ReLU function.

[0037] Further preferably, the activation function of the output of the tensor multilayer perceptron layer is a ReLU function.

[0038] Preferably, according to the present invention, the stochastic gradient descent method is used to optimize the ANCF model, and in each iterative sampling, a stratified sampling rule is formulated to uniformly sample the positive samples and negative samples of the behavioral interactions.

[0039] Further preferably, the stratified sampling rule is:

[0040] For every positive sample of user-item interaction, sample ζ negative samples;

[0041] The sampled positive samples of the interaction include all user-item overall interaction samples and user-item-aspect interaction samples, the number of which is L;

[0042] The ζ+L samples are used as training samples and sent to the ANCF model.

[0043] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of an aspect neural collaborative filtering recommendation method based on Chinese review mining when executing the computer program.

[0044] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of an aspect neural collaborative filtering recommendation method based on Chinese review mining.

[0045] An aspect neural collaborative filtering recommendation system based on Chinese review mining, including aspect term extraction model and aspect neural collaborative filtering model;

[0046] The aspect term extraction model includes corpus preprocessing module, word segmentation module, aspect extraction module, and aspect clustering module;

[0047] Among them, the corpus preprocessing module is used to preprocess the Chinese review dataset; the word segmentation module is used to perform word segmentation on the Chinese review dataset; the aspect extraction module is used to formulate aspect extraction rules for comments based on dependency syntax and part-of-speech rules to extract aspect terms; the aspect clustering module is used to send the extracted aspects into the latent Dirichlet allocation model for aspect clustering and convert them into potential factors;

[0048] The aspect neural collaborative filtering model includes an input module, a generalized tensor decomposition module, a tensor multilayer perceptron module, a neural collaborative filtering module, and an output module;

[0049] Among them, the input module is used to feed the underlying data into the network model; the generalized tensor decomposition module and the multilayer perceptron module are used to extract potential interaction feature tensors from the potential features of users, items, and aspects; the neural collaborative filtering module is used to fit the extracted potential interaction feature tensors with the actual tensor data for training; the output module is used to obtain the prediction score and realize the process of learning the neural collaborative filtering module using the actual interaction tensor.

[0050] The present invention achieves the following beneficial effects through the above solution:

[0051] The present invention implements an aspect neural collaborative filtering recommendation system based on Chinese review mining, which can extract aspects from Chinese reviews and make accurate and efficient recommendations.

[0052] The aspect term extraction model uses dependency syntax to break the reliance on dictionaries, and uses latent Dirichlet distribution to cluster aspects and transform them into potential features. The present invention proposes an unsupervised and efficient aspect extraction method that is suitable for all Chinese reviews, and enables it to be directly combined with the aspect neural collaborative filtering model.

[0053] Compared with the traditional context-based tensor decomposition, the aspect neural collaborative filtering model introduces aspect interaction information, focuses on user behavior from the perspective of potential interaction, and achieves joint attention at the aspect level and item level through stratified sampling. The aspect neural collaborative filtering model increases the interpretability and accuracy of recommendations from a semantic perspective, reduces data complexity and tensor sparsity, solves the cold start problem to a certain extent, increases recommendation coverage, and explores users' potential hobbies.

[0054] Compared with the basic neural collaborative filtering model, third-order tensor decomposition is used instead of matrix decomposition, and stratified sampling is used to achieve joint attention at the aspect level and item level. At the same time, tensor stratified perceptron is used to learn third-order feature interactions, and the fusion training recommendation of aspect collaborative filtering and neural network is realized without increasing the complexity of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of the aspect neural collaborative filtering recommendation method based on Chinese review mining of the present invention.

[0056] Figure 2 This is a schematic diagram of the network structure of the ANCF model of the present invention.

[0057] Figure 3 A graph is constructed for visualizing the potential interaction feature tensor of the present invention.

[0058] Figure 4 This is a visualization example diagram of the stratified sampling rules for training samples of the present invention.

[0059] Figure 5 It is a structural diagram of the aspect neural collaborative filtering recommendation system based on Chinese review mining of the present invention;

[0060] Figure 6 This is an example diagram of aspect extraction in the aspect neural collaborative filtering recommendation method based on Chinese review mining in the present invention. DETAILED DESCRIPTION

[0061] The principles and specific implementations of the present invention are further described below in conjunction with specific examples and accompanying drawings. The examples are only used to explain the present invention but not to limit the scope of the present invention.

[0062] Example 1

[0063] A neural collaborative filtering recommendation method based on Chinese review mining, such as Figure 1 As shown, including:

[0064] Preprocess and segment the Chinese review dataset;

[0065] Formulate aspect extraction rules for comments based on dependency syntax and part-of-speech rules to extract aspect terms;

[0066] After removing stop words from the extracted aspect terms, they are fed into the latent Dirichlet distribution model for aspect clustering and converted into latent factors to construct a Chinese review-aspect dataset;

[0067] Integrate the user, item, and aspect information contained in the Chinese review-aspect dataset to construct a potential interaction feature tensor;

[0068] Send the potential interaction feature tensor to the ANCF model for training to obtain the trained ANCF model;

[0069] The Chinese review-aspect dataset is fed into the trained ANCF model in the form of user id-item id-aspect id to obtain the recommendation results, where the aspect id is 1, representing the overall interaction between the user and the item.

[0070] Example 2

[0071] The difference between the aspect neural collaborative filtering recommendation method based on Chinese review mining described in Example 1 is that:

[0072] The Chinese review dataset was preprocessed, including: deleting comments with poor sentence structure, fake reviews, malicious reviews, and user comments with less than 10 comments. The criteria for comments with poor sentence structure are: they do not conform to Chinese grammar and cannot determine the specific aspects of the comments, such as "Ling Si Si", "Four Stars", "Sa Ge Hao Ge" and other comments; the criteria for user comments with fake reviews and malicious reviews are: if a user has 5 or more identical or similar comments in one day, all his comments will be deleted. Comments from users with less than 10 comments are deleted to prevent problems when dividing the training set and test set.

[0073] This example uses the Douban Chinese movie review dataset, which is publicly available on Kaggle and contains 1,182,419 data items. Each user has at least one review. Examples are shown in Tables 1 and Figure 6 As shown:

[0074] Table 1

[0075]

[0076] In Table 1, “techniques”, “skills”, “fighting”, “sense of rhythm”, and “atmosphere” belong to the same aspect category; “plot” and “setting” belong to the same aspect category; “male and female” and “female lead” belong to the same aspect category; “picture” belongs to one aspect category; and “ending song” belongs to one aspect category.

[0077] According to dependency syntax and part-of-speech rules, the aspect extraction rules of comments are formulated to mine comments, and the extracted aspects are saved as comment aspect word library; including:

[0078] Assume that there are N Chinese reviews, and perform the following steps (1) to (5) for each Chinese review:

[0079] (1) Select the nth Chinese comment and perform word segmentation on it;

[0080] (2) Perform dependency syntactic analysis to obtain all the subjects W pointed to by the subject-verb relationship (SBV) in the Chinese comments. s n , and with the subject W s n Words W in a parallel relationship (COO) c n ;

[0081] (3) Get m dependency syntax extraction words:

[0082] (4) Filter the words in the Chinese comments by their parts of speech and extract all nouns and gerunds Names

[0083] (5) Get the final aspect word

[0084] After removing stop words, the extracted aspects are sent to the latent Dirichlet distribution model for aspect clustering and converted into latent factors, which means: the final aspect words are sent to the latent Dirichlet distribution model for clustering, the number of clusters is set to L, the correlation index λ, and the aspect term group is obtained. After removing stop words, the Chinese review dataset is sent to the latent Dirichlet distribution model to model the topic of the review dataset, and each topic in the latent Dirichlet distribution is regarded as an aspect, where each aspect is composed of a group of aspect terms, and due to the characteristics of the latent Dirichlet distribution, an aspect term can belong to several different aspects with different probabilities; where the number of aspects L∈[4,20], if the number of aspects is too small, it will reduce the supplementary effect of the sparse tensor, if the number of aspects is too large, it will affect the training effect of the ANCF model and it is easy to overfit.

[0085] Integrate the user, item, and aspect information contained in the Chinese review dataset to construct a potential interaction feature tensor, which means: let M, N, and L represent the number of users, items, and aspects, respectively, and transform the one-hot encoded user-item-aspect potential interaction tensor Definition formula (I):

[0086]

[0087] Among them, y ijk is an element of the potential interaction tensor Y, where a value of 1 indicates that there is an interaction between user i and aspect k of item j, and a value of 0 indicates that there is no interaction.

[0088] The aspect dimension is divided into two layers. The first layer takes the real interaction as the overall interaction feature between the user and the item, and the remaining layers insert the user, item, and aspect interaction features, such as Figure 3 As shown, the specific performance is: when k = 1, y ijk The value of indicates whether user i and item j interact or not (1 or 0). When k>1, y ijk The value of indicates whether user i and item j interact with each other in aspect k (1 or 0).

[0089] like Figure 2 As shown, the ANCF model includes an input layer, an embedding layer, a generalized tensor decomposition layer, a tensor multilayer perceptron layer, a neural collaborative filtering layer, and an output layer;

[0090] The hyperparameter settings of the ANCF model in this embodiment are as shown in Table 2:

[0091] Table 2

[0092]

[0093]

[0094] In the input layer, M user IDs, N item IDs, and L aspect IDs are one-hot encoded and mapped to user, item, and aspect feature vectors respectively. And the three sets of feature vectors They are fed into the embedding layer respectively;

[0095] In the embedding layer, the embedding layer is a fully connected layer that constructs the three sets of feature vectors in the input layer. Mapped to a low-dimensional dense feature representation vector with a vector dimension of r, the potential mapping encoding is performed, such as Figure 4 As shown, specifically: The latent factor matrix defining users, items and aspects is expressed as Where r represents the dimension of the latent space, the obtained user potential interaction feature tensor is Item potential interaction feature tensor Aspect latent interaction feature tensor

[0096] In the generalized tensor decomposition layer, the extended generalized tensor decomposition method is used to decompose the potential interaction feature tensor; specifically, as shown in formula (II):

[0097]

[0098] In formula (II), Predicting latent interaction feature tensor elements for generalized tensor decomposition layers; a out and h represent the activation function and edge weight of the output layer respectively, ⊙ represents the element-wise product of the vector;

[0099] In the tensor multilayer perceptron layer, a hidden layer is added to the connection vector inside the perceptron, and a multilayer perceptron that flattens the tensor is used to learn the interaction between the latent features of users, items, and aspects, and to understand the interaction between the latent feature vectors of users, items, and aspects with a nonlinear model;

[0100] The nonlinear model definition is as follows:

[0101]

[0102]

[0103] …

[0104]

[0105]

[0106] Among them, M x ,b x and a x They represent the weight tensor, bias vector, and activation function of the x-th layer perceptron, respectively. is the mapping function of the x-th layer tensor multilayer perceptron, z x is the mapping vector of the xth layer, x = 1...n, n is the number of perceptron layers, σ and h are the activation function and edge weight of the last layer of the output layer respectively; ε MLP Tensor of potential interactions predicted by the multilayer perceptron layer.

[0107] In the neural tensor decomposition layer, the outputs of the tensor multi-layer perception layer and the generalized tensor decomposition layer are concatenated and fed back to the neural collaborative filtering layer to predict the score using the latent vector;

[0108] Specifically, the prediction model is expressed as:

[0109]

[0110] in, Refers to the prediction score of the final output of the ANCF model, V f is the model parameter of the interaction function f; function f is a multi-layer neural network, expressed as:

[0111]

[0112]

[0113] in, and They represent the mapping functions of the output layer and the yth neural collaborative filtering layer, respectively, y = 1...m, m represents the number of layers of the multi-layer neural network;

[0114] In the output layer, the likelihood function is used to constrain the prediction scores to [0,1]:

[0115] Specifically, as follows:

[0116]

[0117] in, is the model likelihood function, represents the set of observed interactions in Y, represents the set of negative instances, all y ijk The values ​​of constitute the set Y, the prediction score represents the possibility that user i is associated with aspect k of item j.

[0118] Using the negative logarithm of the likelihood function as the loss function, namely the binary cross entropy loss, this method pays special attention to the binary attributes of the implicit data:

[0119]

[0120] The final output prediction tensor is represented as follows:

[0121]

[0122]

[0123]

[0124] Among them, ε GTF and ε MLP They represent the prediction tensors of the generalized tensor decomposition and the tensor multilayer perceptron components respectively; σ is the output layer activation function, and the ReLU function is selected; h is the ε calculated by the ANCF model through standard back propagation GTF and ε MLP The weight parameter of .

[0125] The activation function of the output layer uses the ReLU function because it has been shown to be non-saturated, more biologically reasonable, and makes the model less likely to be overfitted.

[0126] The activation function of the tensor multilayer perceptron layer output is the ReLU function.

[0127] The stochastic gradient descent (SGD) method is used to optimize the ANCF model. In each iterative sampling, a stratified sampling rule is formulated to evenly sample the positive and negative samples of behavioral interactions.

[0128] like Figure 4 As shown, the stratified sampling rule is:

[0129] For every positive sample of user-item interaction, sample ζ negative samples;

[0130] The sampled positive samples of the interaction include all user-item overall interaction samples and user-item-aspect interaction samples, the number of which is L;

[0131] The ζ+L samples are used as training samples and sent to the ANCF model.

[0132] Example 3

[0133] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the aspect neural collaborative filtering recommendation method based on Chinese review mining in embodiment 1 or 2 are implemented.

[0134] Example 4

[0135] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the aspect neural collaborative filtering recommendation method based on Chinese review mining in embodiment 1 or 2.

[0136] Example 5

[0137] A neural collaborative filtering recommendation system based on Chinese review mining, such as Figure 5 As shown, it includes the Aspect Term Extraction (ATE) model and the Aspect-based Neural collaborative filtering (ANCF) model;

[0138] The aspect term extraction model includes a corpus preprocessing module, a word segmentation module, an aspect extraction module, and an aspect clustering module; wherein the corpus preprocessing module is used to preprocess the Chinese review dataset; the word segmentation module is used to perform word segmentation on the Chinese review dataset; the aspect extraction module is used to extract aspect terms by formulating aspect extraction rules for comments based on dependency syntax and part-of-speech rules; the aspect clustering module is used to send the extracted aspects into the latent Dirichlet distribution model for aspect clustering and convert them into potential factors;

[0139] The aspect neural collaborative filtering model includes an input module, a generalized tensor decomposition module, a tensor multilayer perceptron module, a neural collaborative filtering module, and an output module; wherein the input module is used to feed the underlying data into the network model; the generalized tensor decomposition module and the multilayer perceptron module are used to extract potential interaction feature tensors from the potential features of users, items, and aspects; the neural collaborative filtering module is used to fit the extracted potential interaction feature tensors with the actual tensor data for training; the output module is used to obtain prediction scores and realize the process of using the actual interaction tensors to learn the neural collaborative filtering module.

Claims

1. A neural collaborative filtering recommendation method based on Chinese review mining, characterized in that: include: Preprocess and segment the Chinese review dataset; Formulate aspect extraction rules for comments based on dependency syntax and part-of-speech rules to extract aspect terms; After removing stop words from the extracted aspect terms, they are fed into the latent Dirichlet distribution model for aspect clustering and converted into latent factors to construct a Chinese review-aspect dataset; Integrate the user, item, and aspect information contained in the Chinese review-aspect dataset to construct a potential interaction feature tensor; Send the potential interaction feature tensor to the ANCF model for training to obtain the trained ANCF model; The Chinese review-aspect dataset is fed into the trained ANCF model in the form of user id-item id-aspect id to obtain the recommendation results, where the aspect id is 1, representing the overall interaction between the user and the item; The ANCF model includes input layer, embedding layer, generalized tensor decomposition layer, tensor multilayer perceptron layer, neural collaborative filtering layer, and output layer; In the input layer, M user IDs, N item IDs, and L aspect IDs are one-hot encoded and mapped to user, item, and aspect feature vectors respectively. And the three sets of feature vectors They are fed into the embedding layer respectively; In the embedding layer, the embedding layer is a fully connected layer that constructs the three sets of feature vectors in the input layer. Mapped to a low-dimensional dense feature representation vector with a vector dimension of r, perform potential mapping encoding, and user potential interaction feature tensor Item potential interaction feature tensor Aspect latent interaction feature tensor In the generalized tensor decomposition layer, the extended generalized tensor decomposition method is used to perform tensor decomposition on the potential interaction feature tensor; In the tensor multilayer perceptron layer, a hidden layer is added to the connection vector inside the perceptron, and a multilayer perceptron that flattens the tensor is used to learn the interaction between the latent features of users, items, and aspects, and to understand the interaction between the latent feature vectors of users, items, and aspects with a nonlinear model; In the neural tensor decomposition layer, the outputs of the tensor multi-layer perception layer and the generalized tensor decomposition layer are concatenated and fed back to the neural collaborative filtering layer to predict the score using the latent vector; In the output layer, a likelihood function is used to constrain the prediction scores to [0,1].

2. According to claim 1, the aspect neural collaborative filtering recommendation method based on Chinese review mining is characterized in that: The Chinese review dataset is preprocessed, including: deleting user comments with poor sentence structure, fake reviews, malicious reviews, and fewer than a certain number of comments.

3. The aspect neural collaborative filtering recommendation method based on Chinese review mining according to claim 1 is characterized in that: According to dependency syntax and part-of-speech rules, the aspect extraction rules of comments are formulated to mine comments, and the extracted aspects are saved as comment aspect word library; including: Assume that there are N Chinese reviews, and perform the following steps (1) to (5) for each Chinese review: (1) Select the nth Chinese comment and perform word segmentation on it; (2) Perform dependency syntactic analysis to obtain all the subjects pointed to by the subject-predicate relationship in the Chinese comments and the subject Words in parallel relationship (3) Get m dependency syntax extraction words: (4) Filter the words in the Chinese comments by their parts of speech and extract all nouns and gerunds Names (5) Get the final aspect word 4. The aspect neural collaborative filtering recommendation method based on Chinese review mining according to claim 1 is characterized in that: After removing stop words, the extracted aspects are sent to the latent Dirichlet distribution model for aspect clustering and converted into latent factors, which means: the final aspect words are sent to the latent Dirichlet distribution model for clustering, the number of clusters is set to L, the correlation index is λ, and the aspect term group is obtained.

5. The aspect neural collaborative filtering recommendation method based on Chinese review mining according to claim 1 is characterized in that: Integrate the user, item, and aspect information contained in the Chinese review dataset to construct a potential interaction feature tensor, which means: let M, N, and L represent the number of users, items, and aspects, respectively, and transform the one-hot encoded user-item-aspect potential interaction tensor Definition formula (I): Among them, y i(k is an element of the potential interaction tensor Y, where a value of 1 indicates that there is an interaction between user i and aspect k of item j, and a value of 0 indicates that there is no interaction.

6. The aspect neural collaborative filtering recommendation method based on Chinese review mining according to claim 1 is characterized in that: The activation function of the output layer is the ReLU function.

7. The aspect neural collaborative filtering recommendation method based on Chinese review mining according to claim 1 is characterized in that: The activation function of the tensor multilayer perceptron layer output is the ReLU function.

8. The aspect neural collaborative filtering recommendation method based on Chinese review mining according to claim 1 is characterized in that: The stochastic gradient descent method is used to optimize the ANCF model. In each iterative sampling, a stratified sampling rule is formulated to evenly sample the positive and negative samples of behavioral interactions.

9. The aspect neural collaborative filtering recommendation method based on Chinese review mining according to claim 8, characterized in that: The stratified sampling rules are: For every positive sample of user-item interaction, sample ζ negative samples; The sampled positive samples of the interaction include all user-item overall interaction samples and user-item-aspect interaction samples, the number of which is L; The ζ+L samples are used as training samples and sent to the ANCF model.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the aspect neural collaborative filtering recommendation method based on Chinese review mining described in any one of claims 1-9 are implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the aspect neural collaborative filtering recommendation method based on Chinese review mining described in any one of claims 1-9 are implemented.

12. An aspect-based neural collaborative filtering recommendation system based on Chinese review mining, characterized in that: Including aspect term extraction model and aspect neural collaborative filtering model; The aspect term extraction model includes corpus preprocessing module, word segmentation module, aspect extraction module, and aspect clustering module; Among them, the corpus preprocessing module is used to preprocess the Chinese review dataset; the word segmentation module is used to perform word segmentation on the Chinese review dataset; the aspect extraction module is used to formulate aspect extraction rules for comments based on dependency syntax and part-of-speech rules to extract aspect terms; the aspect clustering module is used to send the extracted aspects into the latent Dirichlet allocation model for aspect clustering and convert them into potential factors; The aspect neural collaborative filtering model includes an input module, a generalized tensor decomposition module, a tensor multilayer perceptron module, a neural collaborative filtering module, and an output module; Among them, the input module is used to feed the underlying data into the network model; the generalized tensor decomposition module and the multi-layer perceptron module are used to extract the potential interaction feature tensor from the potential features of users, items, and aspects; the neural collaborative filtering module is used to fit the extracted potential interaction feature tensor with the actual tensor data for training; the output module is used to obtain the prediction score and realize the process of using the actual interaction tensor to learn the neural collaborative filtering module; The ANCF model includes input layer, embedding layer, generalized tensor decomposition layer, tensor multilayer perceptron layer, neural collaborative filtering layer, and output layer; In the input layer, M user IDs, N item IDs, and L aspect IDs are one-hot encoded and mapped to user, item, and aspect feature vectors respectively. And the three sets of feature vectors They are fed into the embedding layer respectively; In the embedding layer, the embedding layer is a fully connected layer that constructs the three sets of feature vectors in the input layer. Mapped to a low-dimensional dense feature representation vector with a vector dimension of r, perform potential mapping encoding, and user potential interaction feature tensor Item potential interaction feature tensor Aspect latent interaction feature tensor In the generalized tensor decomposition layer, the extended generalized tensor decomposition method is used to perform tensor decomposition on the potential interaction feature tensor; In the tensor multilayer perceptron layer, a hidden layer is added to the connection vector inside the perceptron, and a multilayer perceptron that flattens the tensor is used to learn the interaction between the latent features of users, items, and aspects, and to understand the interaction between the latent feature vectors of users, items, and aspects with a nonlinear model; In the neural tensor decomposition layer, the outputs of the tensor multi-layer perception layer and the generalized tensor decomposition layer are concatenated and fed back to the neural collaborative filtering layer to predict the score using the latent vector; In the output layer, a likelihood function is used to constrain the prediction scores to [0,1].

Images