A method for implementing a recommendation system based on generating user representations through adversarial learning

Abstract

The application provides a recommendation system implementation method for generating user representation based on adversarial learning, and relates to the technical fields of recommendation system and machine learning. The method first collects user information and converts it into a vector form; then, based on the collected user information, uses a self-encoder module to extract a low-dimensional representation of the user's purchase preference; then generates the user's own preference representation through a conditional perceptual adversarial generation network module; generates a social preference representation through a social perceptual adversarial generation network module; then uses contrast learning technology to learn and train the three preference representations, and distinguish the preference representations of the user and other users from each other; finally, fuses the three preference representations generated by different modules of the same user to obtain more accurate recommendations. The method uses low-dimensional embedding to bridge the distribution gap between user embeddings from different parts, effectively uses the adversarial generation network to generate and improve the potential user embedding to obtain better recommendations.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of recommendation system and machine learning, and particularly relates to a method for implementing a recommendation system based on an adversarial learning to generate user representation. BACKGROUND

[0002] With the development of the information age, the total amount of information on the Internet is rapidly increasing. Pictures, videos, and texts are flooding the user's online space and constantly being updated. In this case, recommendation systems have emerged. Recommendation systems can better solve the problem of information distribution. Through appropriate algorithm design, more matching information can be pushed to the user, increasing the user's experience. Personalized recommendation refers to analyzing and mining user behavior to discover the user's personalized needs and interest characteristics, and recommending information or goods that the user may be interested in to the user. Currently, recommendation systems have become an indispensable part of most Internet platform software systems. In recent years, the rapid development of deep learning has led to the widespread application of deep learning-based recommendation systems in the industry. While greatly improving the recommendation effect, the recommendation system model used has become more complex and difficult to debug, which has led to a huge demand for training data. Under the existing technology, a large number of studies have shown that deep learning methods such as automatic encoders can be used to improve the performance of recommendation systems. However, the model will produce some noise that is inconsistent with the user's preferences during the learning process, and some information will be lost during the learning process. In the field of image generation, generative adversarial networks can be used to generate pictures with some information or make originally blurred pictures clearer. This feature can be used to remove the noise produced by the model and compensate for the model in the recommendation task. More and more algorithms are now tending to supplement the user's feature preferences to achieve more accurate and personalized recommendations, such as some social-based recommendation algorithms based on graph convolution networks, which have been applied to industrial recommendation scenarios.

[0003] Currently, the problem of data sparsity has a significant impact on the performance of such recommendation algorithms, making it difficult to extract accurate user preference information from these data.

[0004] In recent years, in addition to classical methods, deep learning models have also achieved great success. These nonlinear neural networks can automatically learn effective representations from large amounts of data, significantly improving prediction accuracy. In recent years, how to use deep learning models to improve recommendation performance has become a hot topic.

[0005] To overcome the data redundancy problem of recommendation systems, most researchers propose to use denoising auto-encoder (DAE) to learn compact representation from redundant data. Some DAE-based methods focus on how to understand user preferences directly from user rating information through DAE network. Experimental results show that there is a great improvement compared with traditional linear models such as matrix factorization method. At the same time, other works try to use DAE model to learn compact representation from auxiliary information to help improve collaborative filtering-based methods to solve data sparsity problems such as content, label or image.

[0006] In order to better utilize social media, many studies have been proposed to improve the performance of social recommendation using neural network technology. Some works propose to learn user representation from ratings through denoising auto-encoder. Then they propose a new collaborative filtering method to initialize user preferences and measure trust similarity using the learned results. This is an embedding-based recommendation method that utilizes the deep structure of social networks and rating patterns.

[0007] The key problem of these models is that some noise inconsistent with user preferences will be generated in the learning process, and part of the information will be lost in the learning process. In addition, in the generation field, generative adversarial network can be used to generate data with some information, or to make the originally sparse data more abundant. This feature can be used to remove the noise generated by the model and compensate for the model in the recommendation task. SUMMARY

[0008] The technical problem to be solved by the present application is to solve the above-mentioned problems of the prior art, and to provide a recommendation system implementation method based on adversarial learning to generate user representation, which can alleviate the sparsity and unreliability of explicit social relationships and significantly improve the performance of social recommendation.

[0009] To solve the above technical problems, the technical solution adopted by the present application is: a recommendation system implementation method based on adversarial learning to generate user representation,

[0010] Collect user information and convert it into vector form; the user information includes user attribute information, user historical behavior information and user social information; at the same time, construct a social connection graph according to the social information;

[0011] Based on the collected user information, use the auto-encoder module to extract the low-dimensional representation of user purchase preference;

[0012] Generate user's own preference representation through conditional perceptual adversarial generation network module;

[0013] Generate social preference representation through social perceptual adversarial generation network module;

[0014] Use contrastive learning technology to train the three preference representations and distinguish the preference representations of the user from those of other users;

[0015] The three preference representations generated by different modules for the same user are fused to obtain more accurate recommendations.

[0016] Preferably, the autoencoder module uses an autoencoder to extract a low-dimensional representation of the user's purchase preference. en The autoencoder includes an encoder En and a decoder De. The encoder maps a high-resolution user information vector to a low-dimensional user embedding, and the decoder reconstructs the user embedding encoded by the encoder into a high-dimensional user vector to help the autoencoder learn and train.

[0017] Preferably, the condition-aware adversarial generative network module defines a generative model to generate the user's own preference representation The generation model includes a generator and a discriminator; the user's attribute information is encoded into a vector and then used as the conditional embedding c of the generation model. u ; The generator is based on user historical behavior information and conditionally embeds c u As a constraint, generate a vector e representing user preferences r The vector representing the user preference is connected to the low-dimensional representation of the user's purchase preference extracted by the autoencoder module as the input of the discriminator. The discriminator discriminates the user preference extracted by the generator and denoises the low-dimensional representation of the user's purchase preference.

[0018] Preferably, the objective function of the generative model is as follows:

[0019]

[0020] Among them, En(x) represents the low-dimensional representation of user purchase preferences, n represents the number of users, and x i represents the user information vector input to the autoencoder module by user i, z j represents the vector initialized by user j using user social information, c u represents the conditional probability of the user attribute information as the generator to extract the probability distribution, G(·), D(·) represent the generator and discriminator respectively, θ G ,θ D Represent the parameters of the generator and discriminator respectively;

[0021] The parameters of the generator G(·) and the discriminator D(·) are trained and optimized in an alternating manner by solving the minimax game problem:

[0022]

[0023] Preferably, the social-aware generative adversarial network module comprises a social fusion module and a generative adversarial module;

[0024] For each user u, the social information vector of each user is transformed into an initial user preference representation x u by the user's social connection graph u and the corresponding conditional embedding c u As the input of the social fusion module, the user preference representation is diffused by a multi-layer graph convolution network, and the fused social representation e s of the user u is outputted The social fusion module captures the initial social interest of the user from different types of input data;

[0025] The social fusion module in the social-aware generative adversarial network module is modeled as:

[0026]

[0027] wherein, represents the initial social representation of the user u, W is a transformation matrix, and g(·) is a transformation function;

[0028] At each graph convolution layer l+1, the social representation of the user u is recursively outputted by the multi-layer graph convolution network with diffusion operation This iteration step stops when the recursion process reaches a predefined depth;

[0029] For the user u, given its l-th layer social representation The updated user social representation is modeled, and the output of the l+1-th layer of the social fusion module is:

[0030]

[0031] wherein, S u is the friend set of the user u, is the l-th layer user social representation of the user u, is the fused social representation of the user u from its neighbor users in the social connection graph, represents the fused weight between the user and the social friend;

[0032] The updated user social representation of each user is the fusion of the user neighbor information and the user social representation of the user at the previous layer l, and the final output of the social fusion module is wherein, L is the number of layers of the multi-layer graph convolution network set in advance;

[0033] In the process of social fusion, the attention network is used to learn the fusion weight of the l+1 layer of the user u, as shown in the following formula:

[0034]

[0035] Wherein, M represents a multi-layer perceptron;

[0036] The fusion weight is normalized again, as shown in the following formula:

[0037]

[0038] The fusion social representation of each user u is In the input adversarial generation module, the adversarial generation module dynamically uses the layer-by-layer graph convolution to recursively simulate the latent representation of the user social information in the user social network, and obtains the social preference representation es of the user.

[0039] Preferably, the three preference representations are learned and trained using a contrastive learning technique, and the specific method for distinguishing the preference representations of the user and other users from each other is as follows:

[0040] For the same user, the user representations obtained by different processing of the autoencoder module, the conditional perception adversarial generation network module and the social perception adversarial generation network module are used as positive sample pairs, and the user representations of different users are used as negative sample pairs, and the loss function of the infoNCE model is used to make similar user representations closer to each other after training.

[0041] For the same user i, the different representations (e i ,e′ i ) of the same user are regarded as positive pairs, and the representations (e i ,e′ j ) from different users i, j are used as negative sample pairs for contrastive learning; the contrastive loss is defined as follows:

[0042]

[0043] Wherein, τ is a temperature hyperparameter that can be adjusted by contrastive learning, N represents the number of all sample pairs, and s() is a similarity function;

[0044] Preferably, the specific method for obtaining more accurate recommendations by using the three preference representations generated by the fusion same user module is as follows:

[0045] After the reconstruction of user information by the autoencoder module, the enrichment of user information by the conditional perception adversarial generation network module, and the social fusion by the social perception adversarial generation network module, the fusion representation of the user u is divided into three parts, as shown in the following formula:

[0046]

[0047] Among them, R u It means that the user’s fusion representation e is converted by the decoder in the self-encoding module u Convert to recommended results, e u is the result of the fusion of the three parts of user representation, namely Represents an aggregate function.

[0048] Then define the final loss function for user rating as:

[0049]

[0050] where f(·) is a function that measures the difference between the list of inputs and the estimated outputs, R, They represent the actual recommendation results and the predicted recommendation results respectively, and μ is a hyperparameter to avoid overfitting.

[0051] The beneficial effects of adopting the above technical solution are as follows: the present invention provides a method for implementing a recommendation system based on adversarial learning to generate user representations, which uses low-dimensional embedding to bridge the distribution gap between user embeddings from different parts, and effectively utilizes adversarial generative networks to generate and improve potential user embeddings to obtain better recommendations. Compared with social-based recommendation algorithms, the method of the present invention has the advantage of training speed and can greatly save computing resources, especially computing resources on the client side. Due to the introduction of social data, there will not be excessive loss of user preference features, and the theoretical performance upper limit is better than traditional methods. It can better generate enhanced user preference feature information, enrich user features, and enhance the diversity of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of a method for implementing a recommendation system based on adversarial learning to generate user representations provided by an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of the social integration process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0055] In this embodiment, a method for implementing a recommendation system based on adversarial learning to generate user representations is provided. Figure 1As shown, first, user information is collected, including user historical behavior information, such as purchase, like, etc., user attribute information, such as gender, age, etc., and user social information, such as attention and being followed between users, friend status, etc. For each user, these information is collected and converted into vector form; at the same time, a social connection graph is constructed according to the social information; then, based on the collected user information, a self-encoder module is used to extract a low-dimensional representation of the user's purchase preference; a conditional perceptual adversarial generation network module is used to generate the user's own preference representation; a social perceptual adversarial generation network module is used to generate a social preference representation; then, a contrastive learning technique is used to learn and train the three preference representations, and distinguish the preference representations of the user and other users from each other; finally, the three preference representations generated by different modules for the same user are fused to obtain more accurate recommendations.

[0056] Since the adversarial generation model is trained adversarially through minimax game, it is difficult to directly fit the distribution between random noise and original high-dimensional purchase information vector (dimension equal to number of items). Therefore, the present application first uses a self-encoder to extract a low-dimensional purchase preference representation (dimension k), and then through an adversarial generation network, it is easy to fit the distribution between random noise and low dimension.

[0057] Therefore, the self-encoder module uses an auto-encoder to extract a low-dimensional representation e of the user's purchase preference en The auto-encoder includes an encoder En and a decoder De, the encoder is a fully connected network for mapping a high-resolution user information vector to a low-dimensional user embedding, and the decoder reconstructs the user embedding encoded by the encoder into a user high-dimensional vector to help the auto-encoder learn and train. There are two advantages of using an auto-encoder to extract user embeddings from data. First, the auto-encoder extraction effectively preserves high-level features in the main user features to reconstruct the ground truth vector. It is easier to train the generator in the generation model on the extracted features to produce more rich information and improve the hidden embedding. In this sense, the dimension is greatly reduced. Second, the generator only needs to learn how to map from the input noise to the extracted low-dimensional embedding instead of the high-dimensional user vector, which greatly facilitates the training of the generation model.

[0058] Generative adversarial networks (GAN) is a kind of deep learning model, and is one of the most promising methods for unsupervised learning on complex distribution in recent years. The model generates a good output through the mutual game learning of (at least) two models in the framework: the generation model and the discrimination model. In the original model theory, it is not required that both parts of the model are neural networks, but only need to be functions that can fit the corresponding generation and discrimination. But in practice, deep neural networks are generally used as generators and discriminators. An excellent GAN application needs a good training method, otherwise the output may not be ideal due to the freedom of the neural network model.

[0059] While the discriminator restricts the generator, some noise information generated by the generation model from the random vector is removed, while the diversity of the model generated is also retained. We can use this point to remove the noise information that does not conform to the user's preference while retaining some diversified information.

[0060] Therefore, the conditional perceptual adversarial generation network module defines a generation model to generate the user's own preference representation The generation model includes a generator and a discriminator; the attribute information of the user is encoded into a vector and used as a conditional embedding of the generation model

[0061] The generator generates a vector representing the user's preference based on the user's historical behavior information and with the conditional embedding as a constraint; and the vector representing the user's preference is connected with the low-dimensional representation of the user's purchase preference extracted by the autoencoder module as the input of the discriminator. The discriminator discriminates the user's preference extracted by the generator and realizes the denoising processing of the low-dimensional representation of the user's purchase preference.

[0062] The objective function of the generation model is as follows:

[0063]

[0064] Wherein, En(x) represents the low-dimensional representation of the user's purchase preference, n represents the number of users, x i represents the user information vector input to the autoencoder module by the user i, z j represents the vector initialized by the user j using the user social information, c u represents the attribute information of the user as the conditional probability of the probability distribution extracted by the generator, G(·), d(·) respectively represent the generator and the discriminator, θ G , θ D respectively represent the parameters of the generator and the discriminator.

[0065] The parameters of the generator G(·) and the discriminator D(·) are trained and optimized in an alternating manner by solving the minimax game problem:

[0066]

[0067] The social-aware adversarial generation network module includes a social integration module and an adversarial generation module, which are used to achieve social integration of users, such as Figure 2 As shown;

[0068] For each user u, the social information vector of each user is converted into the initial user preference representation x through the user's social connection graph u , the initial user preference is represented by x u and the corresponding conditional embedding c u As the input of the social fusion module, the user preference representation is diffused through the multi-layer graph convolutional network, and the fused social representation e of user u is output s ,The social integration module captures users’ initial social interests from different ,types of input data;

[0069] The social integration module in the social-aware adversarial generation network module is modeled as:

[0070]

[0071] in, represents the initial social representation of user u, W is a transformation matrix, and g(·) is a transformation function;

[0072] In this embodiment, the bias term is omitted when modeling the social integration module.

[0073] Specifically, in each graph convolution layer l+1, the social representation of user u is Taking the previous layer l as input, the updated social representation with diffusion operation is recursively output through the multi-layer graph convolutional network This iterative step stops when the recursive process reaches a predefined depth.

[0074] For user u, given its l-th layer social representation Updated user social media Modeling is performed, and the output of the l+1th layer of the social integration module is:

[0075]

[0076] Among them, S u is the friend set of user u, is the l-th level user social representation of user u, is the fused social representation of user u from its neighbor users in the social connection graph, denotes the fusion weight between user and social friend;

[0077] It can be seen that the user social representation of each user update is the fusion of user neighbor information and user social representation of the previous layer l, and the final output of the social fusion module is where L is the number of layers of the multi-layer graph convolutional network set in advance.

[0078] In the process of social aggregation using the social-aware generative adversarial network module, a simple idea is to use the average pooling operation to aggregate the embeddings from the neighbor users of u; However, it ignores the different interest weights from users, because different users have different importance when representing the item representation. Therefore, it is necessary to use an attention network to learn the fusion weight, so as to better capture the preferences of different users;

[0079] In the above process of social fusion, an attention network is used to learn the fusion weight of the l+1 layer of user u, as shown in the following formula:

[0080]

[0081] where M represents a multi-layer perceptron (MLP);

[0082] The fusion weight is normalized, and an exponential function is used to ensure that each fusion weight is greater than 0, as shown in the following formula:

[0083]

[0084] The fusion social representation of each user u is input into the generative adversarial module, and the generative adversarial module dynamically uses the layer-by-layer graph convolution to recursively simulate the latent representation of the user social information in the user social network, obtaining the social preference representation e of the user s .

[0085] The three preference representations are learned and trained using contrastive learning techniques, and the preference representations of the user and other users are distinguished from each other, so that the user representation is more explicit and the preference is more clear. The specific method is:

[0086] For the same user, the user representations obtained by the autoencoder module, the conditional-aware generative adversarial network module and the social-aware generative adversarial network module are used as positive sample pairs, and the user representations of different users are used as negative sample pairs. Using the loss function of the infoNCE model, similar user representations are closer to each other after training;

[0087] For the same user i, the different representations (e i ,e′i ) are considered as positive pairs, and representations (e i j ) are considered as negative pairs for contrastive learning; the contrastive loss is defined as follows:

[0088]

[0089] where τ is a temperature hyper-parameter for contrastive learning, N represents the number of all sample pairs, and s() is a similarity function.

[0090] After reconstruction of user information by the auto-encoder module, enrichment of user information by the conditional perceptual adversarial generation network module, and social integration by the social perceptual adversarial generation network module, the fused representation of the user u is divided into three parts, as shown in the following formula:

[0091]

[0092] where R u represents the conversion of the fused representation e u of the user into a recommended result by the decoder in the auto-encoder module, e u is the result after fusion of the three-part user representation, i.e. represents an aggregation function, and the concat function is used in this embodiment.

[0093] Then, the final loss function for the user score is defined as:

[0094]

[0095] where f(·) is a function for measuring the list difference between the input and the estimated output, such as the KL(·) function, R,R u represent the actual recommended result and the predicted recommended result (i.e., the recommended result output by the model) respectively; μ is a hyper-parameter for avoiding overfitting. In this embodiment, μ>0 is set, and the positive pairs and the negative pairs are obtained from the same batch of training sets during the training process. The obtained loss is used to optimize the parameters of the overall model using any gradient descent-based algorithm.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present application.​

Claims

1. A method for implementing a recommendation system based on adversarial learning to generate user representations, characterized by: Collect user information and convert it into vector form; the user information includes user attribute information, user historical behavior information and user social information; at the same time, build a social connection graph based on the social information; Based on the collected user information, an autoencoder module is used to extract a low-dimensional representation of the user's purchase preferences; Generate user preference representations through a condition-aware adversarial generation network module; Generate social preference representations through a social-aware generative adversarial network module; Use contrastive learning technology to train the three preference representations and distinguish the preference representations of the user from those of other users; Fusion of three preference representations generated by different modules for the same user to obtain more accurate recommendations; The autoencoder module uses an autoencoder to extract a low-dimensional representation of user purchase preferences ; the autoencoder comprises an encoder and a decoder , the encoder maps a high-resolution user information vector to a low-dimensional user embedding, and the decoder reconstructs the user embedding encoded by the encoder into a high-dimensional user vector to help the autoencoder learn training; The condition-aware generative adversarial network module defines a generative model to generate a user's own preference representation ; the generative model comprises a generator and a discriminator; attribute information of the user is encoded into a vector and then embedded as a condition of the generative model ; Generator based on user historical behavior information and in conditional embedding To constrain, generate a vector representing user preferences ; The vector representing the user preference is connected to the low-dimensional representation of the user's purchase preference extracted by the autoencoder module as the input of the discriminator. The discriminator judges the user preference extracted by the generator and denoises the low-dimensional representation of the user's purchase preference. The objective function of the generative model is as follows: ； wherein, a low-dimensional representation of a user's purchase preference, denotes the number of users, denotes a user a user information vector input to the autoencoder module, denotes a user a vector initialized with user social information, denotes user attribute information as a conditional probability for the generator to extract a probability distribution, denote the generator and the discriminator, respectively, denote the parameters of the generator and the discriminator, respectively; Training the optimization generator in an alternating fashion by solving a minimax game problem and parameters of the discriminator ​ ； The social-aware adversarial generation network module includes a social integration module and an adversarial generation module; For each user , transforming each user’s social information vector into the initial user preference representation through the user’s social connection graph , the initial user preference is represented by and the corresponding conditional embedding As the input of the social integration module, the user preference representation is diffused through the multi-layer graph convolutional network and the user Fusion social representation , is the number of layers of the multi-layer graph convolutional network. The social fusion module captures the user’s initial social interests from different types of input data; The fused social representation is input into the adversarial generation module, which dynamically uses layer-by-layer graph convolution to recursively simulate the potential representation of user social information in the user's social network to obtain the user's social preference representation. . 2.The method of claim 1, wherein: The social fusion module in the social-aware adversarial generative network module is modeled as: ； wherein, represents an initial social representation of a user , is a transformation matrix, is a transformation function; at each graph convolution layer , by taking as input the social representation of the user from the previous layer , the updated social representation with diffusion operation is recursively output by the multi-layer graph convolution network ; this iteration step stops when the recursive process reaches a predefined depth; For a user , given its first layer social representation , the updated user social representation is modeled at the second layer output of the social fusion module as: ； wherein, is a set of friends of the user , is a k-th layer user social representation of the user , is a k-th layer user social representation of the user , is a fused social representation from its neighbor users in the social connection graph for the user denotes a fused weight between the user and the social friend; User social representation updated by each user is the fusion of user neighbor information and user social representation at previous layer The final output of the social fusion module is where, is the number of layers of the multi-layer graph convolution network. 3.The method of claim 2, wherein: The social fusion module uses an attention network to learn the fusion weights of the first layer of the user as shown in the following formula:​ ； wherein represents a multilayer perceptron; Then normalize the fusion weights as shown in the following formula: 。 4.The method of claim 3, wherein: The specific method of using contrastive learning technology to learn and train the three preference representations and distinguish the preference representations of the user and other users is as follows: For the same user, the user representations obtained by different processing of the autoencoder module, the conditional-aware adversarial generative network module, and the social-aware adversarial generative network module are used as positive sample pairs, and the user representations of different users are used as negative sample pairs. The loss function of the infoNCE model is used to make similar user representations closer to each other after training; For the same user different representations of the same user are considered as positive pairs, and representations from different users are considered as negative pairs as negative samples pairs for contrastive learning; the contrastive loss is defined as follows: ； wherein, is a temperature hyperparameter that regulates the contrastive learning, represents the number of all sample pairs, is a similarity function. 5.The method of claim 4, wherein: The specific method of fusing the three preference representations generated by different modules for the same user to obtain more accurate recommendations is as follows: After the reconstruction of user information by the self-encoder module, the enrichment of user information by the conditional generative adversarial network module, and the social integration of the social generative adversarial network module, the fusion representation of the user is divided into three parts, as shown in the following formula: ； wherein, represents the conversion of the user's fused representation by the decoder in the encoding module into a recommended result, is the result after the three-part user representation fusion, i.e. , represents an aggregation function; Then define the final loss function for user rating as: ； wherein, is a function that measures the list difference between the input and the estimated output, respectively represent the actual recommendation result and the predicted recommendation result, is a hyperparameter to avoid overfitting.

Images