A recommendation method and system based on hybrid denoising autoencoder

Abstract

A recommendation method and system based on a hybrid denoising autoencoder, by inputting user-item rating data into a pre-trained generative adversarial network to complete the missing values in the user-item rating data, for each user, the rating values of each item are sorted in descending order to generate a recommendation list. The autoencoder is used as the generator of the generative adversarial network, and the original user-item rating data is added with noise during training, and the user-item rating data with noise is used for training, which improves the robustness of the generative adversarial network. The hybrid use of the autoencoder and the generative adversarial network can achieve complementary advantages and fully exert their respective advantages, learn deeper features of the rating data, and the generated adversarial network obtained after training can fully learn the difference between noisy and denoised data, complete the user-item rating data, and alleviate the data sparsity problem.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, and specifically to a recommendation method and system based on a hybrid denoising autoencoder. Background Technology

[0002] Recommender systems play a crucial role in assisting user decision-making and promoting system service development in various human-computer interaction systems such as e-commerce platforms, music platforms, and video software. A high-performance recommendation algorithm can not only provide users with a more satisfying user experience but also bring significant economic benefits to merchants and service providers. Therefore, improving recommendation accuracy is particularly important. Traditional recommendation systems mostly rely on user-item rating data obtained by users rating various items. However, not every user will rate every item, resulting in a severe data sparsity problem in user-item rating data. How to solve this problem has become a key research focus in the field of recommender systems. Summary of the Invention

[0003] This invention provides a recommendation method and system based on a hybrid denoising autoencoder, used to complete missing values ​​in user-item rating data to alleviate the data sparsity problem.

[0004] According to the first aspect, one embodiment provides a recommendation method based on a hybrid denoising autoencoder, comprising:

[0005] Obtain user-project rating data;

[0006] The user-item rating data is input into a pre-trained generative adversarial network to fill in the missing values ​​in the user-item rating data;

[0007] For each user, their ratings for each item are sorted in descending order to generate a recommendation list;

[0008] Output the recommended list;

[0009] The generative adversarial network (GAN) includes a generator and a discriminator. The generator is an autoencoder, and the GAN is trained in the following manner:

[0010] Obtain raw user-project rating data with complete rating values;

[0011] Noise is added to the original user-item rating data to obtain noisy user-item rating data;

[0012] The original user-item rating data is used as the real sample, and the noisy user-item rating data is used as the noise sample. The network is trained according to a preset objective function to obtain the generative adversarial network.

[0013] In one embodiment, the step of training the generative adversarial network by using the original user-item rating data as real samples and the noisy user-item rating data as noise samples according to a preset objective function includes:

[0014] In the generator training step, the parameters of the discriminator are fixed, and the noisy user-item rating data is input into the generator to obtain denoised user-item rating data X. N The denoised user-project rating data X N The parameters of the generator are updated by backpropagation algorithm according to the preset first objective function input into the discriminator, where N represents the number of layers of the autoencoder;

[0015] The discriminator training step involves fixing the parameters of the generator and inputting the noisy user-item rating data into the generator to obtain denoised user-item rating data X. N The denoised user-project rating data X N The noisy user-item rating data is input into the discriminator, and the parameters of the discriminator are updated through a backpropagation algorithm according to a preset second objective function.

[0016] The generator training step and the discriminator training step are executed alternately until the first objective function and the second objective function converge.

[0017] In one embodiment, the first objective function is:

[0018]

[0019] in

[0020]

[0021] l max =X N -X1,

[0022] D(X N ) represents the denoised user-project rating data X N The result obtained from the discriminator is input, E[·] represents the expected value, and X N ~P noisy (X N ) represents X N It conforms to the distribution of noisy signals, where α and β are preset coefficients, and X1 represents the noisy user-item rating data;

[0023] The second objective function is:

[0024]

[0025] Among them, X1~P noisy (X1) indicates that X1 conforms to a noisy signal distribution, G(X1) represents the result obtained by inputting X1 into the generator, and D(X1,X) represents the result obtained by inputting X1 into the generator. N ) indicates that X1 and X N The result obtained by inputting the discriminator is D(G(X1),X). N ) represents G(X1) and X N The result obtained by inputting the discriminator.

[0026] In one embodiment, the first objective function and the second objective function are determined to have converged when the values ​​of the first objective function and the second objective function are less than a preset convergence threshold.

[0027] In one embodiment, the values ​​of α and β are 0.8 and 0.1, respectively.

[0028] In one embodiment, the noise is additive Gaussian noise.

[0029] In one embodiment, the number of layers of the autoencoder is greater than 3.

[0030] According to a second aspect, one embodiment provides a recommendation system based on a hybrid denoising autoencoder, comprising:

[0031] The data acquisition module is used to acquire user-project rating data;

[0032] A generative adversarial network (GAN) includes a generator and a discriminator. The generator is an autoencoder. The GAN is connected to the data acquisition module and is used to acquire the user-item rating data from the data acquisition module and fill in the missing values ​​in the user-item rating data by calculation.

[0033] The training module is used to train the generative adversarial network to obtain the generative adversarial network;

[0034] The recommendation list generation module is connected to the generative adversarial network and is used to obtain the completed user-item rating data. Then, for each user, the rating values ​​of each item are sorted in descending order to generate a recommendation list.

[0035] An output module, connected to the recommendation list generation module, is used to output the recommendation list;

[0036] The training module includes:

[0037] The training data acquisition unit is used to acquire raw user-item rating data with complete rating values;

[0038] The noise-adding unit is connected to the training data acquisition unit and the generative adversarial network, respectively, and is used to add noise to the original user-item rating data to obtain noisy user-item rating data, and input the noisy user-item rating data into the generative adversarial network for training;

[0039] An optimizer is used to train the generative adversarial network by taking the original user-item rating data as real samples and the noisy user-item rating data as noise samples, according to a preset objective function.

[0040] In one embodiment, the optimizer is trained to obtain the generative adversarial network in the following manner:

[0041] Alternately execute the following generator training steps and discriminator training steps until the preset first and second objective functions converge:

[0042] In the generator training step, the parameters of the discriminator are fixed, and the noisy user-item rating data is input into the generator to obtain denoised user-item rating data X. N The denoised user-project rating data X N The parameters of the generator are updated by backpropagation algorithm according to the first objective function in the discriminator, where N represents the number of layers of the autoencoder;

[0043] The discriminator training step involves fixing the parameters of the generator and inputting the noisy user-item rating data into the generator to obtain denoised user-item rating data X. N The denoised user-project rating data X N The noisy user-item rating data is input into the discriminator, and the parameters of the discriminator are updated through backpropagation algorithm according to the second objective function.

[0044] According to a third aspect, one embodiment provides a computer-readable storage medium, characterized in that the medium stores a program that can be executed by a processor to implement the recommended method as described in the first aspect above.

[0045] According to the recommendation method and system based on hybrid denoising autoencoders in the above embodiments, generative adversarial networks (GANs) are used to complete user-item rating data. During training, the original user-item rating data is denoised, and the denoised user-item rating data is used for training, which improves the robustness of the GAN. In addition, using an autoencoder as the generator of the GAN can fully utilize the characteristic of the autoencoder to reconstruct the original input, further improving the accuracy of the GAN in predicting rating data. Moreover, the hybrid use of the two can complement each other's strengths and weaknesses, learn deeper features of the rating data, and the resulting GAN can fully learn the differences between noisy and denoised data to complete user-item rating data, alleviate the data sparsity problem, and ultimately improve the recommendation efficiency and accuracy of the entire recommendation system. Attached Figure Description

[0046] Figure 1 A flowchart illustrating a recommendation method based on a hybrid denoising autoencoder, according to one embodiment;

[0047] Figure 2 This is a schematic diagram of user-item rating data in one embodiment;

[0048] Figure 3 This is a flowchart illustrating the training process of a generative adversarial network according to one embodiment.

[0049] Figure 4 This is a schematic diagram illustrating the structure and training principle of a generative adversarial network according to one embodiment.

[0050] Figure 5 This is a schematic diagram of the structure of an 8-layer self-encoder according to one embodiment;

[0051] Figure 6 This is a schematic diagram of the structure of a recommendation system based on a hybrid denoising autoencoder, according to one embodiment.

[0052] Figure 7 This is a user interface of a recommendation system based on a hybrid denoising autoencoder, as described in one embodiment.

[0053] Figure 8 User interface 2 for a recommendation system based on a hybrid denoising autoencoder, as described in one embodiment;

[0054] Figure 9 User interface three of a recommendation system based on a hybrid denoising autoencoder, as described in one embodiment;

[0055] Figure 10 User interface four of a recommendation system based on a hybrid denoising autoencoder, as an example;

[0056] Figure 11User interface five of a recommendation system based on a hybrid denoising autoencoder, as described in one embodiment. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0058] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0059] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0060] Big data and artificial intelligence technologies learn the implicit features of data by building models, mapping data from different dimensions to the same space to obtain joint feature representations. In this application, the applicant applies big data and artificial intelligence technologies to traditional recommendation algorithms, which not only solves the problem that traditional recommendation algorithms struggle to obtain deep-level features of users and items, but also obtains joint feature representations of users and items from different dimensions, thus alleviating the data sparsity problem. Specifically, this application uses deep learning to predict missing values ​​in user-item rating data and complete the user-item rating data. Deep learning mainly employs generative adversarial networks (GANs) to improve the predictive ability for missing rating data. To improve the accuracy of predicting missing rating data, the characteristic that autoencoders can reconstruct the original data is utilized, and an autoencoder is introduced as the generator part in the GAN. The combined model can fully learn the differences between noisy and denoised data, generating more robust noise-free rating data, thus achieving the goal of completing the user-item rating data.

[0061] The technical solution of this application is described below. Please refer to... Figure 1 The recommended method based on a hybrid denoising autoencoder in one embodiment of this application includes steps 110 to 140, which are described in detail below.

[0062] Step 110: Obtain user-item rating data. The user-item rating data is in matrix form, storing the rating values ​​of user u for item i, such as... Figure 2 As shown, the rows of the matrix represent users, and the columns represent items, such as movies, music, and products. Taking movies as an example, the user-item rating data stores users' ratings of movies, with ratings ranging from 1 to 5. In reality, since it's impossible for every user to rate every item, the user-item rating data contains many missing values.

[0063] Step 120: Input the user-item rating data into the pre-trained generative adversarial network (GAN) to fill in the missing values ​​in the user-item rating data. The GAN consists of a generator and a discriminator. In this application, an autoencoder is used as the generator. The GAN of this application will be described in detail below.

[0064] Step 130: For each user, sort their ratings for each item in descending order to generate a recommendation list. Items with higher ratings are more likely to match the user's preferences, so prioritize them for recommendation.

[0065] Step 140: Output a recommendation list for users to view, in order to achieve the purpose of recommendation.

[0066] The training process of generative adversarial networks is explained below. Please refer to [link / reference]. Figure 3 In one embodiment, the training process of the generative adversarial network includes steps 121 to 123, which are described in detail below.

[0067] Step 121: Obtain raw user-item rating data with complete rating values ​​for training. MovieLens, a publicly available movie dataset commonly used in recommendation systems, can be used as the raw user-item rating data.

[0068] Step 122: Add noise to the original user-item rating data to obtain noisy user-item rating data. To improve the robustness of the generative adversarial network (GAN), this application adds noise to the user-item rating data used for training, and then trains the GAN, which effectively enhances its robustness. The added noise can be additive Gaussian noise, etc.

[0069] Step 123: Using the original user-item rating data as real samples and the noisy user-item rating data as noise samples, train the generative adversarial network according to the preset objective function to obtain the generative adversarial network.

[0070] Please refer to Figure 4 Generative Adversarial Networks (GANs) typically consist of a generator G and a discriminator D. The generator G reconstructs a "fake sample" from the input noisy samples to simulate the real sample as closely as possible. In this application, the generator G reconstructs the input noisy user-item rating data to obtain denoised user-item rating data, minimizing the error between the reconstructed data and the original data. An autoencoder can achieve this, therefore, it is used as the generator G in this application. The discriminator D judges the difference between the "fake sample" and the real sample. In this application, the discriminator D is used to judge the error between the denoised user-item rating data reconstructed by the autoencoder and the original user-item rating data before noisy input. The discriminator D is similar to a binary classifier, classifying different inputs by compressing and extracting features, and outputting high or low values ​​accordingly. The discriminator D's task is to minimize the error between the denoised data and the original data while reducing its own output loss and improving its error judgment ability. The loss of the generator G and the discriminator D can be measured by a pre-set objective function, and their parameters can be updated to achieve the training effect.

[0071] Traditional autoencoders typically have only three layers, resulting in low data utilization efficiency, an inability to extract deeper user and item features, and problems such as insufficient interpretability of recommendations and weak model representation capabilities. Although traditional autoencoder models can handle high-dimensional and sparse raw data, their scalability as the model needs to incorporate additional data sources will be reduced to some extent without optimization.

[0072] Therefore, this application proposes an innovative improvement to the traditional autoencoder in one embodiment. The traditional autoencoder is a three-layer network structure, typically consisting of an encoding layer, a hidden layer, and a decoding layer. To improve the generalization ability of the traditional autoencoder, the depth of the autoencoder is increased. Furthermore, since training with noisy data gives the autoencoder denoising capabilities, the traditional shallow autoencoder network is improved into a deep denoising autoencoder. This results in more robust data representation and the learning of deeper features, thus improving generalization ability. Combining the autoencoder with a generative adversarial network can further improve the prediction ability for missing rating data, effectively addressing the data sparsity problem in traditional recommendation algorithms.

[0073] In specific implementations, the number of layers in the autoencoder can be increased, for example, by setting it to 8 layers, 10 layers, etc. The following description uses an 8-layer autoencoder as an example; its structure is as follows: Figure 5 As shown, Figure 5 In this model, X0 represents the original user-item rating data, and X1 represents the noisy user-item rating data. The autoencoder consists of an encoding part and a decoding part. The encoding part takes the noisy, high-dimensional, sparse noisy user-item rating data X1 as input, performs dimensionality reduction and compression operations through a hidden layer, and calculates the latent feature vectors P for users and items. u and q i This process can be represented by the following formula:

[0074]

[0075] Where R (u) and R (i) These are the user and project data from the original user-project rating data, respectively. u1 e u2 e u3 e u4 e represents the user feature vectors of each layer in the encoding part. i1 e i2 e i3 e i4 For the feature vectors of each layer of the encoding part, This is the user weight matrix for each layer of the encoding part. For the item weight matrix of each layer of the encoding part, b u1 b u2 b u3 b u4 b represents the user offset for each layer of the encoding section. i1 b i2 b i3 b i4The offset of each layer in the encoding part is f(·), which is a non-linear activation function. Scaled exponential linear units (SeLU) can be selected as the activation function, which has a normalization function.

[0076] The decoding part obtains low-dimensional user-item rating data X4 after dimensionality reduction and compression in the hidden layer y. This low-dimensional data is then reconstructed through a four-layer decoder to produce output data X8, similar to the original user-item rating data—that is, denoised user-item rating data. This stage utilizes the latent feature vectors P of users and items. u and q i This involves reconstructing the input to complete the missing rating data. This process can be represented by the following formula:

[0077]

[0078] The meanings of each parameter are similar to those in the encoding section, and will not be repeated here.

[0079] During training, the generator G and discriminator D are trained alternately. The entire training process mainly includes generator training steps and discriminator training steps. In the generator training step, the parameters of the discriminator D are fixed, and the noisy user-item rating data is input into the generator G. After encoding and decoding by an autoencoder, more robust denoised user-item rating data X is generated. N This stage can be understood as generating generator loss, and then denoising the user-item rating data X. N In the input discriminator D, the parameters of the generator G are updated through the backpropagation algorithm according to the preset first objective function, where N represents the number of layers of the autoencoder.

[0080] In the discriminator training step, the parameters of the generator G are fixed, and the noisy user-item rating data is input into the generator G to obtain the denoised user-item rating data X. N , denoising user-project rating data X N The user-project rating data with added noise is input into the discriminator D to obtain the output value of the discriminator D. This stage can be understood as generating the discriminator loss. Then, according to the preset second objective function, the parameters of the discriminator are updated through the backpropagation algorithm.

[0081] The generator training step and the discriminator training step are executed alternately until the first objective function and the second objective function converge. The applicant has improved the objective function of the traditional generative adversarial network, and the improved first objective function is as follows:

[0082]

[0083] in

[0084]

[0085] l max =X N -X1,

[0086] D(X N () indicates that the noise-reduced user-project rating data X N The result obtained from the input discriminator D, E[·] represents the expected value, X N ~P noisy (X N ) represents X N The signal conforms to a noisy signal distribution, where α and β are preset coefficients, and in one embodiment, they can be 0.8 and 0.1 respectively. Here, a distance function l is introduced. dist and maximum local error l max This improves the accuracy of objective function optimization.

[0087] The second objective function is as follows:

[0088]

[0089] Among them, X1~P noisy (X1) indicates that X1 conforms to a noisy signal distribution, G(X1) represents the result obtained by inputting X1 into the generator G, and D(X1,X) represents the signal distribution. N ) indicates that X1 and X N The result obtained from the input discriminator is D(G(X1),X). N ) represents G(X1) and X N The result obtained from the input discriminator D.

[0090] By training the generator G and the discriminator D alternately, the generator G can continuously learn the differences between the input noisy scoring data and the denoised scoring data, thereby improving its denoising ability. When the values ​​of the first objective function and the second objective function are less than the preset convergence threshold, the first objective function and the second objective function can be considered to have converged. At this time, the generator G has been trained to have a good ability to denoise noisy data.

[0091] During training, after the generator G updates its parameters using the backpropagation algorithm, the generated denoised user-item rating data causes the discriminator D to output high values. The generator G's parameters remain unchanged as the discriminator D continues to be trained, improving its ability to distinguish between real and fake samples. This causes the discriminator D's output value for the denoised user-item rating data generated by the generator G to decrease again. Through repeated iterations, the user-item rating data generated by the generator G becomes closer to clean, noise-free user-item rating data, and its distribution closely resembles the true distribution. Ultimately, this restores the original user-item rating data, completing the entire denoising process. The generated user-item rating data at this point exhibits strong robustness and can fill in missing values ​​in the input user-item rating data, essentially predicting the missing rating values ​​and effectively mitigating the data sparsity problem.

[0092] Based on the aforementioned recommendation method based on a hybrid denoising autoencoder, this application also provides a recommendation system based on a hybrid denoising autoencoder. Please refer to [link / reference]. Figure 6 In one embodiment, the system includes a data acquisition module 1, a generative adversarial network 2, a training module 3, a recommendation list generation module 4, and an output module 5, which are described below.

[0093] Data acquisition module 1 is used to acquire user-project rating data.

[0094] Generative Adversarial Network 2 (GAN2) comprises a generator G and a discriminator D. The generator G is an autoencoder. GAN2 is connected to the data acquisition module 1 and is used to obtain user-item rating data from the data acquisition module 1, and to complete the missing values ​​in the user-item rating data through calculation. As mentioned above, the autoencoder acting as the generator can be an autoencoder with increased layers; the number of layers is not limited to 3, but can be 8, 10, etc. For the specific structure of GAN2, please refer to the above; it will not be repeated here.

[0095] Training module 3 is used to train the generative adversarial network 2. For example... Figure 6 As shown, training module 3 includes a training data acquisition unit 31, a noise-adding unit 32, and an optimizer 33. The training data acquisition unit 31 acquires raw user-item rating data with complete rating values ​​for training. The movie dataset movielens, commonly used in recommendation systems, can be used as the raw user-item rating data.

[0096] The noise-adding unit 32 is connected to the training data acquisition unit 31 and the generative adversarial network 2, respectively. It is used to add noise to the original user-item rating data to obtain noisy user-item rating data, and input the noisy user-item rating data into the generative adversarial network 2 for training. The added noise can be additive Gaussian noise, etc.

[0097] Optimizer 33 is used to train the original user-item rating data as real samples and the noisy user-item rating data as noise samples according to a preset objective function to obtain the generative adversarial network 2.

[0098] During training, the generator G and discriminator D are trained alternately. The entire training process mainly includes generator training steps and discriminator training steps. In the generator training step, the parameters of the discriminator D are fixed, and the noisy user-item rating data is input into the generator G. After encoding and decoding by an autoencoder, more robust denoised user-item rating data X is generated. N This stage can be understood as generating generator loss, and then denoising the user-item rating data X. N In the input discriminator D, the parameters of the generator G are updated through the backpropagation algorithm according to the preset first objective function, where N represents the number of layers of the autoencoder.

[0099] In the discriminator training step, the parameters of the generator G are fixed, and the noisy user-item rating data is input into the generator G to obtain the denoised user-item rating data X. N , denoising user-project rating data X N The user-project rating data with added noise is input into the discriminator D to obtain the output value of the discriminator D. This stage can be understood as generating the discriminator loss. Then, according to the preset second objective function, the parameters of the discriminator are updated through the backpropagation algorithm.

[0100] Optimizer 33 alternately executes the generator training steps and discriminator training steps described above until the first objective function and the second objective function converge. The applicant has improved the objective function of the traditional generative adversarial network, and the improved first objective function is as follows:

[0101]

[0102] in

[0103]

[0104] l max =X N -X1,

[0105] D(X N () indicates that the noise-reduced user-project rating data X N The result obtained from the input discriminator D, E[·] represents the expected value, X N ~P noisy (X N ) represents X NThe signal conforms to a noisy signal distribution, where α and β are preset coefficients, and in one embodiment, they can be 0.8 and 0.1 respectively. X1 represents the noisy user-item rating data. Here, a distance function l is introduced. dist and maximum local error l max This improves the accuracy of objective function optimization.

[0106] The second objective function is as follows:

[0107]

[0108] Among them, X1~P noisy (X1) indicates that X1 conforms to a noisy signal distribution, G(X1) represents the result obtained by inputting X1 into the generator G, and D(X1,X) represents the signal distribution. N ) indicates that X1 and X N The result obtained from the input discriminator is D(G(X1),X). N ) represents G(X1) and X N The result obtained from the input discriminator D.

[0109] By training the generator G and the discriminator D alternately, the generator G can continuously learn the differences between the input noisy scoring data and the denoised scoring data, thereby improving its denoising ability. When the values ​​of the first objective function and the second objective function are less than the preset convergence threshold, the first objective function and the second objective function can be considered to have converged. At this time, the generator G has been trained to have a good ability to denoise noisy data.

[0110] During training, after the generator G updates its parameters using the backpropagation algorithm, the generated denoised user-item rating data causes the discriminator D to output high values. The generator G's parameters remain unchanged as the discriminator D continues to be trained, improving its ability to distinguish between real and fake samples. This causes the discriminator D's output value for the denoised user-item rating data generated by the generator G to decrease again. Through repeated iterations, the user-item rating data generated by the generator G becomes closer to clean, noise-free user-item rating data, and its distribution closely resembles the true distribution. Ultimately, this restores the original user-item rating data, completing the entire denoising process. The generated user-item rating data at this point exhibits strong robustness and can fill in missing values ​​in the input user-item rating data, essentially predicting the missing rating values ​​and effectively mitigating the data sparsity problem.

[0111] The recommendation list generation module 4 is connected to the generative adversarial network 2 to obtain the completed user-item rating data. Then, for each user, the rating values ​​of each item are sorted in descending order to generate a recommendation list.

[0112] Output module 5 is connected to recommendation list generation module 4 and is used to output a recommendation list for users to view, so as to achieve the purpose of recommendation.

[0113] The recommendation method and system based on a hybrid denoising autoencoder, as described in the above embodiments, utilize generative adversarial networks (GANs) to complete user-item rating data. During training, the original user-item rating data is noise-added, and the noisy user-item rating data is used for training, improving the robustness of the GAN. Furthermore, using an autoencoder as the generator for the GAN fully leverages its ability to reconstruct the original input, further improving the accuracy of the GAN's rating data prediction. The combined use of both methods allows for the leveraging of their respective strengths, learning deeper features of the rating data. In some embodiments, the number of autoencoder layers can be increased to form a deep autoencoder, enhancing the generalization ability of the GAN. The trained GAN can effectively learn the differences between noisy and denoised data, completing the user-item rating data, alleviating data sparsity issues, and ultimately improving the overall recommendation efficiency and accuracy of the recommendation system.

[0114] This application applies big data and artificial intelligence technologies to the traditional recommendation field and makes innovative improvements to autoencoders and generative adversarial networks. It can not only effectively make up for the limitations of traditional recommendation algorithms and provide a good supplement and improvement to traditional recommendation algorithms, but also greatly improve the overall performance and scalability of traditional recommendation algorithms. It has certain reference value and practical significance for promoting the development of recommendation algorithms.

[0115] Unlike popular recommendations and ranking system recommendations, the recommendation method and system in this application offer personalized and intelligent recommendations, tailored to each user. This makes the recommendations more interpretable and credible. The difference lies in the fact that these two types of recommendations are not sensitive to personalization and intelligence; they simply present currently popular content to users, which cannot satisfy everyone's tastes. The quality of recommendations hides significant commercial value. The recommendation method and system in this application predict and provide recommendations based on users' existing rating data, achieving a tailored recommendation effect. This makes it easier to provide users with explanations for the recommendations, thereby increasing user trust and reliance on the recommendation system. This tailored recommendation strategy not only brings significant advertising and commercial revenue to service providers but also provides a superior user experience for each user.

[0116] To verify the practicality and effectiveness of the technical solution presented in this application, the applicant selected the MovieLens dataset as the dataset for the specific implementation and applied the technical solution to a movie recommendation system. To facilitate adjustments to computational power, a multi-threaded processing mechanism was incorporated, and the root mean square error (RMSE) was used as a criterion for judging recommendation accuracy. The main function of this recommendation system is to predict a user's rating for a movie—that is, how many points a user might rate a movie if they have watched it—and then recommend movies to the user in descending order based on the predicted ratings. The system allows users to view their personal information and the ratings of movies they have watched.

[0117] The system interface is as follows Figure 7 As shown, you can select the number of threads; here, 8 threads are used for calculation. After clicking "Start," the score calculation will begin. Clicking the "Root Mean Square Error" button will display the predicted root mean square error. The score will be displayed in... Figure 9 The recommendation information interface shown allows users to enter a user ID to receive movie recommendations, sorted in descending order of rating. The ratings are calculated using the recommendation method described in this application. For example, entering "24" will display the predicted rating score that user would give the movie if they had watched it. Movies are recommended to that user based on their predicted rating, from highest to lowest. Selecting the "HaveCommented" option will display details of all movies that the target user has previously watched and rated. For example, for user 24, the details of their rated movies are as follows... Figure 10 As shown. Selecting the UserInfo option allows you to view the user's personal information, such as... Figure 11 As shown.

[0118] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0119] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A recommendation method based on a hybrid denoising autoencoder, characterized in that, include: Obtain user-project rating data; The user-item rating data is input into a pre-trained generative adversarial network to fill in the missing values ​​in the user-item rating data; For each user, their ratings for each item are sorted in descending order to generate a recommendation list; Output the recommended list; The generative adversarial network (GAN) includes a generator and a discriminator. The generator is an autoencoder, and the GAN is trained in the following manner: Obtain raw user-project rating data with complete rating values; Noise is added to the original user-item rating data to obtain noisy user-item rating data; The generative adversarial network (GAN) is trained by using the original user-item rating data as real samples and the noisy user-item rating data as noise samples, according to a preset objective function. The network includes: In the generator training step, the parameters of the discriminator are fixed, and the noisy user-item rating data is input into the generator to obtain denoised user-item rating data. X N The denoised user-project rating data X N The parameters of the generator are updated using a backpropagation algorithm based on a preset first objective function input into the discriminator. N This indicates the number of layers in the autoencoder; The discriminator training step involves fixing the parameters of the generator and inputting the noisy user-item rating data into the generator to obtain denoised user-item rating data. X N The denoised user-project rating data X N The noisy user-item rating data is input into the discriminator, and the parameters of the discriminator are updated through a backpropagation algorithm according to a preset second objective function. The generator training step and the discriminator training step are executed alternately until the first objective function and the second objective function converge; The first objective function is: ， in ， ， This indicates that the denoised user-project rating data will be used. X N Input the result obtained from the discriminator. Indicates the expected value. express X N Conforms to the distribution of noisy signals. α and β The preset coefficients, X 1 represents the noisy user-item rating data; The second objective function is: ， in, express X 1. Conforms to the distribution of noisy signals. Indicates will X 1. Input the result obtained by the generator. Indicates will X 1 and X N The result obtained by inputting the discriminator Indicates will and X N The result obtained by inputting the discriminator.

2. The recommended method as described in claim 1, characterized in that, When the values ​​of the first objective function and the second objective function are less than the preset convergence threshold, the first objective function and the second objective function are determined to have converged.

3. The recommended method as described in claim 1, characterized in that, α and β The values ​​are 0.8 and 0.1, respectively.

4. The recommended method as described in claim 1, characterized in that, The noise is additive Gaussian noise.

5. The recommended method as described in any one of claims 1 to 4, characterized in that, The number of layers in the autoencoder is greater than 3.

6. A recommendation system based on a hybrid denoising autoencoder, characterized in that, include: The data acquisition module is used to acquire user-project rating data; A generative adversarial network (GAN) includes a generator and a discriminator. The generator is an autoencoder. The GAN is connected to the data acquisition module and is used to acquire the user-item rating data from the data acquisition module and fill in the missing values ​​in the user-item rating data by calculation. The training module is used to train the generative adversarial network to obtain the generative adversarial network; The recommendation list generation module is connected to the generative adversarial network and is used to obtain the completed user-item rating data. Then, for each user, the rating values ​​of each item are sorted in descending order to generate a recommendation list. An output module, connected to the recommendation list generation module, is used to output the recommendation list; The training module includes: The training data acquisition unit is used to acquire raw user-item rating data with complete rating values; The noise-adding unit is connected to the training data acquisition unit and the generative adversarial network, respectively, and is used to add noise to the original user-item rating data to obtain noisy user-item rating data, and input the noisy user-item rating data into the generative adversarial network for training; An optimizer is used to train the generative adversarial network by taking the original user-item rating data as real samples and the noisy user-item rating data as noise samples, according to a preset objective function. The optimizer is trained to obtain the generative adversarial network in the following manner: Alternately execute the following generator training steps and discriminator training steps until the preset first and second objective functions converge: In the generator training step, the parameters of the discriminator are fixed, and the noisy user-item rating data is input into the generator to obtain denoised user-item rating data. X N The denoised user-project rating data X N The parameters of the generator are updated using a backpropagation algorithm based on the first objective function input into the discriminator, wherein... N This indicates the number of layers in the autoencoder; The discriminator training step involves fixing the parameters of the generator and inputting the noisy user-item rating data into the generator to obtain denoised user-item rating data. X N The denoised user-project rating data X N The noisy user-item rating data is input into the discriminator, and the parameters of the discriminator are updated through backpropagation algorithm according to the second objective function; The first objective function is: ， in ， ， This indicates that the denoised user-project rating data will be used. X N Input the result obtained from the discriminator. Indicates the expected value. express X N Conforms to the distribution of noisy signals. α and β The preset coefficients, X 1 represents the noisy user-item rating data; The second objective function is: ， in, express X 1. Conforms to the distribution of noisy signals. Indicates will X 1. Input the result obtained by the generator. Indicates will X 1 and X N The result obtained by inputting the discriminator indicates that... and X N The result obtained by inputting the discriminator.

7. A computer-readable storage medium, characterized in that, The medium stores a program that can be executed by a processor to implement the recommended method as described in any one of claims 1-5.

Images