A movie recommendation method based on triple autoencoder combined with a knowledge graph

Abstract

The application discloses a movie recommendation method based on triple auto-encoder combined with a knowledge graph, which comprises the following steps: 1) encoding the comment information between the user and the movie into an emotional classification as the input of an automatic encoder; 2) merging the rating of the movie, auxiliary information and generated comment representation into a semi-automatic encoder for reconstructing the output, learning the low-dimensional feature representation of the extended information through the semi-automatic encoder, fusing the obtained low-dimensional feature representation into the original feature space of the movie, and inputting the new feature as additional information into the semi-automatic encoder model; 3) designing a serial connection of the semi-automatic encoder and the automatic encoder, and comparing the output prediction score matrix of the third auto-encoder with the original score matrix. The application can expand the features of the movie information by using the interaction information between the user and the movie and the knowledge graph, and process the expanded features through the auto-encoder, so that the more accurate recommendation for the user is achieved.

Description

Technical Field

[0001] This invention relates to the field of personalized data recommendation research, and in particular to a movie recommendation method based on a triple autoencoder combined with a knowledge graph. Background Technology

[0002] Nowadays, in all information-related application scenarios, with the rapid development of the Internet, there is an urgent need for personalized recommendations to solve the problem of information overload.

[0003] It is worth noting that many successful recommender systems share various aspects of feature representation learning and have been widely applied to many online services, such as e-commerce and social networks. Existing recommender system methods can be broadly classified into three categories: content-based recommendation, collaborative filtering (CF), and hybrid methods.

[0004] Content-based recommendation methods learn descriptive features of items, calculate the similarity between new items and items liked by users based on these features, and generate the final recommendation. However, this requires well-structured feature content, and users' preferences must be expressible in the form of content features, without explicitly revealing other users' judgments. Collaborative filtering methods discover user preferences and generate recommendations by considering users' historical behavior, but they face problems such as sparsity of the rating matrix and weak generalization ability. To address these issues, matrix factorization techniques have been proposed. The main operation is to learn latent features of users or items from the rating matrix to optimize recommendation accuracy, and it has achieved significant success in practical applications. However, matrix factorization methods have limitations in feature representation learning capabilities because most methods learn user and item feature representations by directly factoring the rating matrix without considering additional information. Hybrid recommendation methods combine multiple methods and attempt to combine their advantages, but it is difficult to obtain auxiliary information from other information sources, and most of the auxiliary information itself is sparse, making it difficult to effectively and directly integrate it into the recommendation system to improve recommendation accuracy. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a movie recommendation method based on a triple autoencoder combined with a knowledge graph. This method can utilize the interaction information between users and movies and the knowledge graph to expand the features of movie information, and connects the semi-autoencoder and the autoencoder in series to make more accurate recommendations for users.

[0006] The objective of this invention is achieved as follows: a movie recommendation method based on a triple autoencoder combined with a knowledge graph, comprising the following steps:

[0007] 1) Encode the comment information between users and movies into sentiment categories, and use this as input to an autoencoder to generate auxiliary information for the project;

[0008] 2) The movie's rating, auxiliary information, and generated review representations are incorporated into a semi-autoencoder for output reconstruction. The semi-autoencoder learns low-dimensional feature representations to expand the information, which are then fused into the movie's original feature space. These new features are treated as additional information. The input is fed into the semi-autoencoder model;

[0009] 3) Design a serial connection between a semi-automatic encoder and an automatic encoder. Obtain the output generated by the semi-automatic encoder and then re-input the generated output into the third automatic encoder. Compare the output prediction rating matrix of the third automatic encoder with the original rating matrix to calculate the prediction accuracy and make more accurate movie recommendations.

[0010] Furthermore, step 1) specifically includes:

[0011] Step 1.1) Use BeautifulSoup web crawling technology to search for comment information of the corresponding movies in the MovieLens dataset;

[0012] Step 1.2) Parse the statement "'#main>section>div.lister>div.lister-list>div>div.review-container>div.lister-item-content>div.content>div.text.show-more__control'" containing the comments to be crawled, filter the search results, extract the required comment information, and use the results as the initial feature expansion.

[0013] Furthermore, step 2) specifically includes:

[0014] Step 2.1) Train the autoencoder model using formulas (1) and (2), retaining the model parameters for the comment information:

[0015] ξ s =f(W s C+b s (1)

[0016] s=g(W′ s ·ξ s +b′ s (2)

[0017] Among them, W s ∈R k1×n and W′ s ∈R n×k1 It is the weight matrix of the autoencoder, b s ∈Rk1×1 ,b′ s ∈R n×1 is the bias vector, f and g are non-linear activation functions, k1 is the feature dimension of the hidden layer unit, s′ is the low-dimensional representation of the hidden features of the first autoencoder; and these are merged into the second semi-autoencoder to sample different subsets from all inputs to capture different representations and reconstruct them;

[0018] Step 2.2) After obtaining the reconstructed review features, a semi-automatic encoder is applied, referencing the rating vector r of movie i. i Other auxiliary information a i Reconstructing comment features i and extended features l i Connect the four vectors to get con(r) i ,a i ,s i ,l i ):

[0019] con(r i ,a i ,s i ,l i = concatenation of r i ,a,s i and l i (3)

[0020] The combined vector is then used to merge and concatenate all the movie information to obtain the input of the semi-automatic encoder as con(R). I A I ,S I ,L I ):

[0021] Step 2.3) The con(R) obtained from step 2.2) I A I ,S I ,L I The input data is fed into the semi-autoencoder model to obtain the compressed reconstructed output; the coding level of the semi-autoencoder is defined as shown in formula (4):

[0022] R′ semi =g(W′f(Wcon(R) I A I ,S I ,L I (4)

[0023] in, and W′∈R k×m Represents the weight matrix; and is the bias vector, and f and g are non-linear activation functions.

[0024] Furthermore, step 3) includes:

[0025] Step 3.1) Design a third autoencoder model to learn the reconstruction of the entire input. The encoding and decoding stages of the triple autoencoder are represented as shown in equations (5) and (6), respectively:

[0026] θ=f(W t R′ semi +b t (5)

[0027] R′=g(W′ t θ+b′ t (6)

[0028] Weight matrix W t and W′ t The 2-norm regularization is added to the objective function, which is expressed as shown in equation (7):

[0029]

[0030] Therefore, the final objective function of the triple autoencoder is shown in equation (8):

[0031] J item =||(R′-R′) semi )|| 2 +αJ r (8)

[0032] Using the L-FBGS algorithm, W is calculated iteratively. t ,W′ t ,b t ,b′ t After the iterative calculation is completed, the reconstruction matrix R′ of the entire input is obtained. Based on R′, items with a user score of more than 0.5 are recommended to the user.

[0033] The present invention adopts the above technical solution, and compared with the prior art, the beneficial effects are as follows: 1) The autoencoder model is used to learn the feature representation of the user's interaction information with the movie. The autoencoder model has the characteristics of fast convergence speed, no need for labels and good effectiveness, which makes the method more practical.

[0034] 2) This invention uses knowledge graphs to extend the features of movie information. At the same time, the additional features obtained by extension are represented by low-dimensional features extracted by an autoencoder. The original user movie rating matrix, the original features, and the extended features processed by the autoencoder are extracted again by a semi-autoencoder. The output is then reconstructed and input into the autoencoder, so that it can be more conveniently and flexibly applied to recommendation models.

[0035] 3) This invention designs a serial connection between a semi-automatic encoder and an automatic encoder to learn better feature representations, thereby improving the accuracy of model recommendations. Attached Figure Description

[0036] Figure 1 Overall framework diagram of the present invention.

[0037] Figure 2 A schematic diagram of the self-encoder model structure in this invention.

[0038] Figure 3 A schematic diagram of the semi-autoencoder model structure in this invention. Detailed Implementation

[0039] like Figure 1 The movie recommendation method shown includes the following steps: (The method is based on a triple autoencoder combined with a knowledge graph.)

[0040] 1) Encode the comment information between users and movies into sentiment categories, and use this as input to an autoencoder to generate auxiliary information for the project;

[0041] Step 1.1) Use BeautifulSoup web crawling technology to search for review information of the corresponding movies in the MovieLens dataset; "https: / / www.imdb.com / title / tt'+value+' / reviews?ref_=tt_urv", where the value parameter in the URL is set to the imdbID of the movie;

[0042] Step 1.2) Parse the statement "'#main>section>div.lister>div.lister-list>div>div.review-container>div.lister-item-content>div.content>div.text.show-more__control'" containing the comments to be crawled, filter the search results, extract the required comment information, and use the results as the initial feature expansion.

[0043] 2) The movie's rating, auxiliary information, and generated review representations are incorporated into a semi-autoencoder for output reconstruction. The semi-autoencoder learns low-dimensional feature representations to expand the information, which are then fused into the movie's original feature space. These new features are treated as additional information. Inputting the data into a semi-autoencoder model enables more accurate movie recommendations;

[0044] Step 2.1) Train the autoencoder model using formulas (1) and (2), retaining the model parameters for the comment information:

[0045] ξ s =f(W s C+b s (1)

[0046] s=g(W′ s ·ξ s +b′ s (2)

[0047] Among them, W s ∈R k1×n and W′ s ∈R n×k1 It is the weight matrix of the autoencoder, b s ∈R k1×1 ,b′ s ∈R n×1 is the bias vector, f and g are non-linear activation functions, k1 is the feature dimension of the hidden layer unit, s′ is the low-dimensional representation of the hidden features of the first autoencoder; and these are merged into the second semi-autoencoder to sample different subsets from all inputs to capture different representations and reconstruct them;

[0048] Step 2.2) After obtaining the reconstructed review features, a semi-automatic encoder is applied, referencing the rating vector r of movie i. i Other auxiliary information a i Reconstructing comment features i and extended features l i Connect the four vectors to get con(r) i ,a i ,s i ,l i ):

[0049] con(r i ,a i ,s i ,l i = concatenation of r i ,a,s i and l i(3)

[0050] The combined vector is then used to merge and concatenate all the movie information to obtain the input of the semi-automatic encoder as con(R). I A I ,S I ,L I );

[0051] Step 2.3) The cat(R) obtained from step 2.2) I A I ,S I ,L I The input data is fed into the semi-autoencoder model to obtain the compressed reconstructed output; the coding level of the semi-autoencoder is defined as shown in formula (4):

[0052] R′ semi =g(W′f(Wcon(R) I A I ,S I ,L I (4)

[0053] in, and W′∈R k×m Represents the weight matrix; and is the bias vector, and f and g are non-linear activation functions.

[0054] 3) Design a serial connection between a semi-automatic encoder and an autoencoder. By acquiring the output generated by the semi-automatic encoder, the output is then re-inputted into a third autoencoder to learn more abstract and higher-level feature representations for personalized recommendations. The output of the third autoencoder (predicted rating matrix) is compared with the original rating matrix to calculate the prediction accuracy, which greatly improves the accuracy of movie recommendations.

[0055] Step 3.1) Design a third autoencoder model to learn the reconstruction of the entire input. The encoding and decoding stages of the triple autoencoder are represented as shown in equations (5) and (6), respectively:

[0056] θ=f(W t R′ semi +b t (5)

[0057] R′=g(W′ t θ+b′ t (6)

[0058] Weight matrix W t and W′ tThe 2-norm regularization is added to the objective function, which is expressed as shown in equation (7):

[0059]

[0060] Therefore, the final objective function of the triple autoencoder is shown in equation (8):

[0061] J item =||(R′-R′) semi )|| 2 +αJ r (8)

[0062] Using the L-FBGS algorithm, W is calculated iteratively. t ,W′ t ,b t ,b′ t After the iterative calculation is completed, the reconstruction matrix R′ of the entire input is obtained. Based on R′, items with a user score of more than 0.5 are recommended to the user.

[0063] This invention can be further illustrated by the following experiments:

[0064] To test the effectiveness of this invention, prediction results were implemented on the MovieLens 100K and MovieLens 1M datasets, respectively. The MovieLens 100K dataset includes 100,000 ratings from 943 users on 1,682 movies, and the MovieLens 1M dataset includes 1,000,209 ratings from 6,040 users on 3,706 movies. The evaluation metric used was the root mean square error (RMSE), calculated as follows: the smaller this value, the better the recommendation system.

[0065]

[0066] Where, r u,i and Let |TestSet| represent the original and reconstructed user u ratings for movie i, respectively, and |TestSet| represent the entire test set.

[0067] To demonstrate the performance of the test results, the 100K dataset used Non-negative Matrix Factorization (NMF), Improved Singular Value Decomposition (SVD++), and a knowledge graph combined with an autoencoder recommender system (PRKG) for comparison; the 1M dataset used Non-negative Matrix Factorization (NMF), a multilayer perceptron-based recommender system (NCF), and a semi-autoencoder-based recommender system (HCRSA) for comparison. The prediction results of the datasets are shown in Table 1. As can be seen from Table 1, the root square error (RMSE) of the prediction results of this invention is better than other methods on the three datasets.

[0068] Table 1. Experimental results of the RMSE index

[0069]

[0070] This invention utilizes user-movie interaction information and knowledge graphs to expand movie features. Then, an autoencoder is used to reduce the dimensionality of the obtained sparse feature information, extracting efficient feature representations, thus solving the problems of insufficient and sparse auxiliary information in recommendation systems. Next, a semi-autoencoder is used to reduce the dimensionality of the original and expanded feature information, extracting features and outputting prediction results. Finally, the prediction results are further compressed and extracted using an autoencoder. The serial connection of the semi-autoencoder and autoencoder learns more abstract and higher-level feature representations for personalized recommendations, achieving more accurate recommendations for users.

[0071] This invention is not limited to the above embodiments. Based on the technical solutions disclosed in this invention, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.

Claims

1. A movie recommendation method based on triple autoencoder combined with a knowledge graph, characterized in that, The steps include: 1) encode the review information between the user and the movie as an emotional classification as an input of an automatic encoder for generating auxiliary information of the item; Step 1.1) search the review information of the corresponding movie in the MovieLens dataset using the BeautifulSoup web crawler technology; Step 1.2) parse the sentence where the crawled review is located, that is, "'#main>section>div.lister>div.lister-list>div>div.review-container>div.lister-item-content>div.content>div.text.show-more__control'", filter the searched web results, and extract the required review information, and the results are used as initial feature expansion; 2) merge the rating of the movie, the auxiliary information and the generated review representation into a semi-automatic encoder for reconstruction output, obtain a low-dimensional feature representation of the expanded information through semi-automatic encoder learning, fuse the obtained low-dimensional feature representation into the original feature space of the movie, and input the new features as additional information into the semi-automatic encoder model; Step 2.1) complete the training of the autoencoder model through formula (1) and formula (2) to reserve the review information model parameters: ξ s = f(W s C+b s ) (1) s = g(W'x + b') (1) s ·ξ s +b′ s ) (2) where W s ∈ R k1×n and W' s ∈ R n×k1 are weight matrices of the autoencoders, b s ∈ R k1×1 and b' s ∈ R n×1 are bias vectors, f and g are non-linear activation functions, and k1 is the feature dimension of the hidden layer units; and merging ξ s into a second semi-automated encoder for sampling different subsets from all inputs to capture different representations and reconstruct review features; Step 2.2) After obtaining the reconstructed review features, apply a semi-automatic encoder, referencing the rating vector r of the movie i i and other auxiliary information a i , the reconstructed review features s i and the extended features l i , concatenate the four vectors to get con(r i , a i , s i , l i ): con(r i ,a i ,s i ,l i )=concatenation of r i ,a,s i and l i (3) By the obtained combination vector, all the movie information is combined and connected to obtain the input of the semi-automatic encoder as con(R I ,A I ,S I ,L I ); Step 2.3) The input data of con(R I I I I ) obtained by step 2.2) is input into the semi- autoencoder model to obtain the compressed reconstruction output; the encoding level of the semi- autoencoder is defined as shown in equation (4):​​​ R' semi = g(W'f(Wcon(R I , A I , S I , L I ) + b) + b1) (4) wherein, and W' ∈ R k×m represents a weight matrix; and is a bias vector, f and g are non-linear activation functions; 3) design a series connection of the semi-automatic encoder and the automatic encoder, obtain the output generated by the semi-automatic encoder, re-input the generated output into a third automatic encoder, compare the output prediction score matrix of the third automatic encoder with the original score matrix, calculate the prediction accuracy, and make more accurate movie recommendations.

2. The movie recommendation method based on triple autoencoder combined with knowledge graph according to claim 1, characterized in that, The step 3) includes: Step 3.1) design a third automatic encoder model to learn the reconstruction of the entire input, and the encoding and decoding stages of the triple autoencoder are represented by formula (5) and (6) respectively: θ = f(W t R′ semi +b t ) (5) R' = g(W') t θ + b' t ) (6) weight matrix W t and W t ′ 2-norm regularization of W is added to the objective function, which is expressed as shown in Equation (7): Therefore, the final objective function of the triple autoencoder is shown in formula (8): J item = ||(R' - R' + αJ semi )| 2 + αJ r (8) Using L-FBGS algorithm, iteratively calculate W t ,W′ t ,b t ,b′ t After the iterative calculation is completed, the reconstruction matrix R' of the entire input is obtained, and according to R', the items with scores exceeding 0.5 on the user are recommended to the user.

Images