An object recommendation method and system based on embedded noise and deep and shallow layer comparison

Abstract

The present invention relates to the technical field of data prediction and recommendation, and in particular to an item recommendation method and system based on embedded noise and deep and shallow comparison. The method comprises obtaining user information and item information of a target user; performing matrix decomposition on the obtained user information and item information to obtain a user matrix and an item matrix as shallow information for comparative learning; performing graph neural network forward propagation operation on the user matrix and the item matrix to obtain user-item interaction information and item-user interaction information as deep information for comparative learning; fusing the shallow information for comparative learning with the deep information for comparative learning to obtain user comprehensive information and item comprehensive information; the present invention retains the global information of the interaction between the user and the item and weakens the local information through the shallow part of comparative learning based on matrix decomposition, thereby retaining the global auxiliary information reflecting the user's demand for the item.

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction and recommendation, and in particular to an item recommendation method and system based on embedded noise and deep and shallow layer comparison. Background Art

[0002] In today's information age, recommendation systems face the challenge of how to accurately recommend content from massive amounts of data. Existing recommendation technologies usually use the user's historical data, that is, the items that the user has interacted with, to discover items that belong to the same field or label as the item and recommend them to the user. The advantage of this recommendation method is that the recommended items are rich and fast, but the large amount of recommended data sometimes dazzles the user and requires more time to select the items they like. Similarly, the quality of recommended items will also vary, resulting in a decrease in user experience; the recommendation method based on ratings can be used as a good way to solve this problem, reduce low-quality problems and thus improve user experience. This method usually recommends high-rated items to users based on the user's historical evaluation (such as ratings), but there are also certain problems. It may be difficult to meet the changing preferences of users, and the reduction of recommended content may affect user choices, thereby reducing user experience. Summary of the invention

[0003] In order to solve the above-mentioned problems, the present invention provides an item recommendation method and system based on embedded noise and deep and shallow layer comparison.

[0004] In a first aspect, the present invention provides an item recommendation method based on embedded noise and deep and shallow layer comparison, which adopts the following technical solution:

[0005] An object recommendation method based on embedding noise and deep and shallow layer comparison, comprising:

[0006] Obtain user information and project information of the target user;

[0007] Perform matrix decomposition on the acquired user information and project information to obtain user matrix and project matrix as shallow information for comparative learning;

[0008] Perform graph neural network forward propagation operations on the user matrix and the item matrix to obtain user-item interaction information and item-user interaction information as deep information for comparative learning;

[0009] The shallow contrastive learning information and the deep contrastive learning information are integrated to obtain comprehensive user information and comprehensive project information.

[0010] After nonlinear processing of user comprehensive information and item comprehensive information, the recommended value of the target item is obtained;

[0011] Based on the set threshold, determine whether the target item is pushed to the target user.

[0012] Furthermore, the acquisition of user information and item information of the target user includes collecting the side relationships of items that have interacted with the target user and related demand information based on the user-item interaction platform to form target user relationship interaction information.

[0013] Furthermore, the obtained user information and project information are subjected to matrix decomposition to obtain a user matrix and a project matrix, including forming a user-project interaction set and a project-user interaction set based on the obtained user information and project information, and performing dimensionality reduction decomposition based on the initial user-item interaction matrix to decompose it into a user matrix and a project matrix, that is, comparative learning of shallow information.

[0014] Furthermore, the user matrix and the project matrix are subjected to a graph neural network forward propagation operation to obtain user-project interaction information and project-user interaction information, including performing a graph neural network forward propagation operation based on the user matrix and the project matrix and adding a mixed noise of uniform noise and Gaussian noise for perturbation in the process; after the operation is completed, the results of each layer are aggregated to form the user's features and the project's features, generating a main view of contrastive learning, i.e., contrastive learning deep information.

[0015] Furthermore, the contrastive learning shallow information and the contrastive learning deep information are fused to obtain user comprehensive information and project comprehensive information, including combining the contrastive learning shallow information and the contrastive learning deep information to obtain a perturbation view of the fused global information, wherein user features are generated by performing matrix multiplication between user-project interaction information and the user matrix and the project matrix generated by matrix decomposition; and project features are generated by performing matrix multiplication between project-user interaction information and the user matrix and the project matrix generated by matrix decomposition.

[0016] Furthermore, after the nonlinear processing of the user comprehensive information and the item comprehensive information, the recommendation value of the target item is obtained, including nonlinear calculation of the main view and the perturbation view, wherein the user positive and negative sample scores and the item positive and negative sample scores are calculated respectively, and the negative sample score is: for each user, the scores of all items are calculated to obtain the logarithmic probability of the negative sample; the positive sample score is: for each user and item pair, the dot product of their embedding vectors is calculated, and then scaled, and the result is limited to a certain range through the clamp function; finally, the positive sample score is added to the negative sample score to obtain the final recommendation value.

[0017] Furthermore, judging whether the target item is pushed to the target user based on the set threshold includes recommending the target item to the user according to a calculation formula of the recommendation value when the recommendation value of the target item is not less than 0.5, otherwise not recommending the target item. The calculation formula of the recommendation value is: , S is the recommended value of the target item.

[0018] In the second aspect, an item recommendation system based on embedding noise and deep and shallow layer comparison includes:

[0019] The data acquisition module is configured to acquire user information and project information of a target user;

[0020] The decomposition module is configured to perform matrix decomposition on the acquired user information and project information to obtain a user matrix and a project matrix as shallow information for comparative learning;

[0021] The interaction module is configured to perform a graph neural network forward propagation operation on the user matrix and the item matrix to obtain user-item interaction information and item-user interaction information as deep information for comparative learning;

[0022] A fusion module is configured to fuse the contrastive learning shallow information and the contrastive learning deep information to obtain user comprehensive information and project comprehensive information;

[0023] The linear module is configured to obtain the recommendation value of the target item after performing nonlinear processing on the user comprehensive information and the item comprehensive information;

[0024] The push module is configured to determine whether the target item is pushed to the target user based on a set threshold.

[0025] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing the object recommendation method based on embedded noise and deep and shallow layer comparison.

[0026] In a fourth aspect, the present invention provides a terminal device including a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being suitable for being loaded by the processor and executing the method for recommending items based on embedded noise and deep and shallow layer comparison.

[0027] In summary, the present invention has the following beneficial technical effects:

[0028] The present invention retains the global information of user-item interaction and weakens local information through the shallow part of contrastive learning based on matrix decomposition, thereby retaining the global auxiliary information reflecting the user's demand for the project; then, the graph neural network forward propagation of the initial user-item interaction matrix is ​​processed through the deep part of contrastive learning and mixed noise interference is added in the process, and user features and project features are obtained after aggregation; then, the global auxiliary information is fused with the user features and project features respectively to obtain the user features and project features fused with the global interaction information; finally, the user features and project features are respectively calculated with the user features fused with the global interaction information and the project features fused with the global interaction information for recommendation value. When the recommendation value reaches the threshold, the project item is recommended to the user, which improves the accuracy of the item recommendation, thereby enhancing the user experience and promoting consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 4 is a schematic diagram of an item recommendation method based on embedded noise and deep and shallow layer comparison according to Example 1 of the present invention. DETAILED DESCRIPTION

[0030] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0031] Example 1

[0032] Reference Figure 1 , an item recommendation method based on embedded noise and deep-shallow layer comparison in this embodiment includes:

[0033] Obtain user information and project information of the target user;

[0034] Perform matrix decomposition on the acquired user information and project information to obtain user matrix and project matrix as shallow information for comparative learning;

[0035] Perform graph neural network forward propagation operations on the user matrix and the item matrix to obtain user-item interaction information and item-user interaction information as deep information for comparative learning;

[0036] The shallow contrastive learning information and the deep contrastive learning information are integrated to obtain comprehensive user information and comprehensive project information.

[0037] After nonlinear processing of user comprehensive information and item comprehensive information, the recommended value of the target item is obtained;

[0038] Based on the set threshold, determine whether the target item is pushed to the target user.

[0039] Specifically, the following steps are included:

[0040] S1. In the user-item interaction formed on user-item interaction platforms (such as JD.com, Taobao, Pinduoduo, NetEase Cloud Music, QQ Music and other user service platforms), based on the items that have interacted with the target user, the side relationships that can reflect the relevant demand information raised by the user when using the items in the past are collected to form the target user relationship interaction information, so as to facilitate the subsequent recommendation of suitable items for the target user based on the comprehensive information characteristics of the target user. The project information is the initial features of the project, including project id, rating, etc., which is a low-dimensional embedding and represents the implicit features of the item; the project information is generated through platform data initialization, and the users who have interacted with the item are collected as side relationships as part of the project information. In addition, it also contains information such as project id and project rating.

[0041] First, based on the data of the user-item interaction platform, all items that have interacted with the target user are counted to form a user-item interaction set and an item-user interaction set.

[0042] S2. Based on the initial user-item interaction matrix, matrix decomposition is performed to decompose it into a user matrix and an item matrix as global auxiliary information to prepare for the subsequent fusion of global information, that is, the shallow part of contrastive learning; the user-item interaction matrix is ​​a sparse matrix created from the index data of users and items. The elements of the sparse matrix are initialized to 1 to represent the interaction between users and items. In order to improve the expressive power of the matrix and avoid the deviation introduced by the different number of interactions between users and items, the matrix is ​​normalized. The specific steps include: first, the total number of interactions between each user and each item is calculated. The total number of user interactions is obtained by summing each row of the matrix, and the total number of item interactions is obtained by summing each column of the matrix. Then, each non-zero element in the matrix is ​​normalized. For each non-zero value in the matrix, it is divided by the product of the square root of the total number of interactions between the user and the item.

[0043] Among them, matrix decomposition is based on the initial user-item interaction matrix, and the matrix is ​​decomposed into a user implicit information matrix and an item implicit information matrix by reducing its dimensionality, thereby improving the accuracy and efficiency of the calculation. This is the existing technology.

[0044] The matrix decomposition operation can be achieved by the following formula:

[0045]

[0046] R is the rating matrix with dimension (n, m), U is the user feature matrix with dimension (n, k), and V is the item feature matrix with dimension (k, m), where k is the rank, which is usually small.

[0047] S3. Based on the initial user-item interaction matrix, perform the forward propagation operation of the graph neural network and add a mixed noise of uniform noise and Gaussian noise for perturbation in the process; after the operation is completed, aggregate the results of each layer to form the user's features and the project's features, and generate the main view of contrastive learning, that is, the deep part of contrastive learning; fuse the global auxiliary information generated by the shallow part with the user feature matrix and the project feature matrix generated by the deep part to generate a perturbation view of contrastive learning that fuses the global information.

[0048] First, user features and item features are obtained through the forward propagation process of the graph neural network. The features can reflect the target user’s interest in the target item. Then, in this process, a mixture of uniform noise and Gaussian noise is added to perturb the user features and item features. The purpose is to improve the uniformity of the features, thereby improving the anti-interference ability and improving the recommendation performance.

[0049] The generation of noise can be achieved by the following formula:

[0050]

[0051] is Gaussian noise with mean 0 and variance 1. is a uniform noise with a value between 0 and 1, l1 and l2 are the noise scaling factors that control the noise ratio.

[0052] The forward propagation process of the graph neural network can be implemented by the following formula:

[0053]

[0054] is the matrix multiplication operation, is the noise scaling factor, which controls the size of the noise.

[0055] Then, the results of the forward propagation of each layer are aggregated. The specific operation of aggregation is addition, and the formula is , E is the user feature or item feature.

[0056] S4. In order to combine the deep and shallow information to obtain the perturbation view of the fused global information, the user latent feature matrix and the project latent feature matrix generated by matrix decomposition are fused with the user feature matrix and the project feature matrix respectively.

[0057] The fusion processing operation can be: performing matrix multiplication of the user matrix and the project matrix generated by matrix decomposition through user-project interaction information to generate user features; performing matrix multiplication of the user matrix and the project matrix generated by matrix decomposition through project-user interaction information to generate project features.

[0058] This can be achieved by the following formula:

[0059]

[0060]

[0061] and They are the user feature matrix that integrates the global feature information and the item feature matrix that integrates the global feature information; and They are the user latent feature matrix and the item latent feature matrix generated by matrix decomposition; and They are user feature matrix and item feature matrix respectively.

[0062] S5. The main view and the perturbation view are subjected to nonlinear operation to obtain a recommendation value of the target item, and it is determined whether the recommendation value is less than a recommendation value threshold; if so, the target item is not recommended to the target user; if not, the target item is recommended to the target user.

[0063] First, after being processed by S1 and S2, the main view and the perturbation view are generated, and the recommendation value is calculated by nonlinear operations on the main view and the perturbation view. The specific operation of calculating the recommendation value is: the main view and the perturbation view respectively calculate the user positive and negative sample scores and the item positive and negative sample scores. The negative sample score is: for each user, the scores of all items are calculated to obtain the logarithmic probability of the negative sample; the positive sample score is: for each user and item pair, the dot product of their embedding vectors is calculated, and then scaled, and the result is limited to a certain range through the clamp function; finally, the positive sample score and the negative sample score are added to the final recommendation value.

[0064] The formula for calculating the negative sample score is as follows:

[0065]

[0066] is the feature vector of the target user that incorporates the global feature information. is the embedding of all items that the target user has interacted with, is a constant with a value of 1e-8, and T is the temperature parameter used for scaling; is the feature vector of the target project that incorporates global feature information. are the embeddings of all users who have interacted with the target item.

[0067] The formula for calculating the positive sample score is as follows:

[0068]

[0069] The function is to fix the calculated value between (-5,5). is the feature vector of the target user that incorporates the global feature information. is user embedded, It is the feature vector of the target object that integrates the global feature information. It is project embedded.

[0070] S6. Finally, the recommended value is calculated as: , S is the recommendation value of the target item. When the recommendation value of the target item is not less than 0.5, the target item is recommended to the user, otherwise it is not recommended.

[0071] Example 2

[0072] This embodiment provides an item recommendation system based on embedded noise and deep-shallow layer comparison, including:

[0073] The data acquisition module is configured to acquire user information and project information of a target user;

[0074] The decomposition module is configured to perform matrix decomposition on the acquired user information and project information to obtain a user matrix and a project matrix as shallow information for comparative learning;

[0075] The interaction module is configured to perform a graph neural network forward propagation operation on the user matrix and the item matrix to obtain user-item interaction information and item-user interaction information as deep information for comparative learning;

[0076] A fusion module is configured to fuse the contrastive learning shallow information and the contrastive learning deep information to obtain user comprehensive information and project comprehensive information;

[0077] The linear module is configured to obtain the recommendation value of the target item after performing nonlinear processing on the user comprehensive information and the item comprehensive information;

[0078] The push module is configured to determine whether the target item is pushed to the target user based on a set threshold.

[0079] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing the object recommendation method based on embedded noise and deep and shallow layer comparison.

[0080] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the method for recommending items based on embedded noise and deep and shallow layer comparison.

[0081] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An object recommendation method based on embedded noise and deep and shallow layer comparison, characterized in that: include: Obtain user information and project information of the target user; Perform matrix decomposition on the acquired user information and project information to obtain user matrix and project matrix as shallow information for comparative learning; Perform graph neural network forward propagation operations on the user matrix and the item matrix to obtain user-item interaction information and item-user interaction information as deep information for comparative learning; The shallow contrastive learning information and the deep contrastive learning information are integrated to obtain comprehensive user information and comprehensive project information. After nonlinear processing of user comprehensive information and item comprehensive information, the recommended value of the target item is obtained; Based on the set threshold, determine whether the target item is pushed to the target user; The obtaining of user information and item information of the target user includes collecting the side relationships of items that have interacted with the target user and related demand information based on the user-item interaction platform to form target user relationship interaction information; The step of performing matrix decomposition on the acquired user information and item information to obtain a user matrix and an item matrix includes forming a user-item interaction set and an item-user interaction set based on the acquired user information and item information, and performing dimensionality reduction decomposition based on the initial user-item interaction matrix to decompose it into a user matrix and an item matrix, i.e., comparative learning of shallow information; The user matrix and the project matrix are subjected to a graph neural network forward propagation operation to obtain user-project interaction information and project-user interaction information, including performing a graph neural network forward propagation operation based on the user matrix and the project matrix and adding a mixed noise of uniform noise and Gaussian noise for disturbance in the process; after the operation is completed, the results of each layer are aggregated to form user features and project features, and a main view of contrastive learning is generated, that is, contrastive learning deep information; The method comprises fusing the shallow contrast learning information and the deep contrast learning information to obtain the user comprehensive information and the project comprehensive information, including combining the shallow contrast learning information and the deep contrast learning information to obtain the perturbation view of the fused global information, wherein the user features are generated by performing matrix multiplication between the user-project interaction information and the user matrix and the project matrix generated by matrix decomposition; and the project features are generated by performing matrix multiplication between the project-user interaction information and the user matrix and the project matrix generated by matrix decomposition; After the user comprehensive information and the item comprehensive information are processed nonlinearly, the recommendation value of the target item is obtained, including nonlinear calculation of the main view and the perturbation view, wherein the user positive and negative sample scores and the item positive and negative sample scores are calculated respectively, and the negative sample score is: for each user, the scores of all items are calculated to obtain the logarithmic probability of the negative sample; the positive sample score is: for each user and item pair, the dot product of their embedding vectors is calculated, and then scaled, and the result is limited to a certain range by the clamp function; finally, the positive sample score and the negative sample score are added to obtain the final recommendation value; The method of judging whether the target item is pushed to the target user based on the set threshold includes: when the target item's recommendation value is not less than 0.5, the target item is recommended to the user; otherwise, the target item is not recommended. The calculation formula of the recommendation value is: , S is the recommended value of the target item.

2. An item recommendation system based on embedded noise and deep-shallow layer comparison, executing the method as claimed in claim 1, characterized in that: include: The data acquisition module is configured to acquire user information and project information of a target user; The decomposition module is configured to perform matrix decomposition on the acquired user information and project information to obtain a user matrix and a project matrix as shallow information for comparative learning; The interaction module is configured to perform a graph neural network forward propagation operation on the user matrix and the item matrix to obtain user-item interaction information and item-user interaction information as deep information for comparative learning; A fusion module is configured to fuse the contrastive learning shallow information and the contrastive learning deep information to obtain user comprehensive information and project comprehensive information; The linear module is configured to obtain the recommendation value of the target item after performing nonlinear processing on the user comprehensive information and the item comprehensive information; The push module is configured to determine whether the target item is pushed to the target user based on a set threshold.

3. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

4. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method as claimed in claim 1 .

Images