Graph collaborative filtering recommendation method based on high-order similarity denoising

Abstract

The invention provides a graph collaborative filtering recommendation method based on high-order similarity denoising. The graph collaborative filtering recommendation method comprises the following steps of obtaining interaction historical data of a user and an article and performing preprocessing; constructing a user-article interaction bipartite graph, estimating second-order similarity in the graph, and performing neighborhood denoising; initial feature vectors are constructed for all users and articles and are initialized; constructing a recommended main task model by adopting an alternating graph convolutional network; constructing a self-supervision auxiliary task, and realizing neighborhood preservation based on second-order similarity; the recommendation main task and the self-supervised learning auxiliary task are combined, a unified optimization target is determined, and multi-task learning training is adopted to obtain feature vectors of the user and the article; and according to the learned feature vectors, predicting scores of non-interactive articles by the target user, and sorting to obtain an article recommendation list. According to the method, through alternate graph convolution based on similarity denoising and self-supervised neighborhood preservation, robust graph collaborative filtering is realized, and the sensitivity of an existing recommendation method to noise data is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation systems, and in particular is a graph collaborative filtering recommendation method based on high-order similarity denoising. Background Art

[0002] Recommendation system is an important technology to solve the problem of information overload. It helps users obtain information efficiently by mining data related to user behavior, thereby achieving the goal of providing personalized services to the public. It has been widely used in scenarios such as online shopping, online learning, and health consultation. Collaborative filtering, as a key component in the recommendation system, aims to obtain users' interest in new items from historical interaction data between users and items (such as historical ratings, browsing history, purchase lists, etc.). Influenced by the Netflix competition, collaborative filtering recommendation methods based on representation learning, represented by matrix factorization (MF), have become mainstream. Its goal is to learn the representation of users and items from user-item interaction data.

[0003] In recent years, as graph neural networks have achieved outstanding results in representation learning, collaborative filtering models based on graph representation learning have received widespread attention. Unlike the traditional method of obtaining representations using only the numbers of users and items (such as MF and other models), since user-item interaction data can be represented in the form of a bipartite graph, by applying graph neural networks (such as NGCF, LightGCN and other models) on the user-item bipartite graph, the collaborative filtering recommendation method based on graph representation learning, namely graph collaborative filtering, can make full use of the high-order similarity in the graph to obtain the representation of users and items, thus significantly improving the recommendation effect.

[0004] However, the interaction data in the recommendation system is often noisy, for example, a user clicks on an item by accident rather than out of interest. Compared with traditional representation learning-based methods, graph-based collaborative filtering recommendation methods are often more sensitive to noisy data. The reason is that the negative impact of noisy interactions is continuously amplified under the message passing mechanism of graph neural networks. Therefore, how to improve the robustness of collaborative filtering recommendation methods based on graph representation learning is a major challenge facing current graph collaborative filtering recommendation methods.

[0005] At present, there are two main ways to implement robust graph collaborative filtering: based on self-supervised learning and based on interactive denoising. The graph collaborative filtering recommendation method based on self-supervised learning introduces additional self-supervised learning tasks on the basis of supervised learning collaborative filtering tasks, so as to capture self-supervised signals based on graph enhancement, representation enhancement or neighborhood enhancement (such as SGL, SimGCL and other models). However, this multi-task joint training method does not fundamentally solve the negative impact of noise interactions existing in the original collaborative filtering tasks and their diffusion in the graph. The graph collaborative filtering recommendation method based on interactive denoising adopts denoising strategies to reduce the impact of unreliable interactions. For example, the RGCF and RocSE models remove the noise edges representing unreliable interactions by introducing graph structure denoising modules. However, it is very difficult to judge the reliability of interactions with limited information. Therefore, interactive denoising has limited effect on the improvement of the robustness of graph collaborative filtering, and more effective denoising methods need to be studied. Summary of the invention

[0006] In view of the problems existing in the above-mentioned prior art, the present invention provides a graph collaborative filtering recommendation method based on high-order similarity denoising. The method considers high-order similarity from the perspective of high-order similarity and obtains a more effective denoising method. It also combines self-supervised learning technology to make full use of reliable second-order similarity. On this basis, a robust graph collaborative filtering method is provided to achieve better recommendation indicators on multiple public data sets, especially higher indicator improvements on sparse data sets.

[0007] In order to achieve the above object, the present invention provides a graph collaborative filtering recommendation method based on high-order similarity denoising, comprising the following steps:

[0008] Step 1: Data preprocessing;

[0009] Obtaining massive historical interaction data between users and items through data processing equipment, and preprocessing the historical interaction data using the data processing equipment to obtain a user-item interaction matrix;

[0010] Step 2: Similarity denoising;

[0011] Construct a bipartite graph of user-item interactions, perform second-order similarity estimation based on the bipartite graph, and perform neighborhood denoising on this basis to obtain the final second-order similarity;

[0012] Step 3: Initialize feature vector;

[0013] Construct and initialize initial feature vectors for all users and items;

[0014] Step 4: Recommend the construction of the main task model;

[0015] The alternating graph convolutional network is used to obtain the final feature vectors of users and items, and the recommendation loss is defined based on this;

[0016] Step 5: Self-supervised auxiliary task construction;

[0017] Define self-supervised loss to achieve neighborhood preservation based on second-order similarity;

[0018] Step 6: Multi-task learning;

[0019] Combine the recommendation main task and the self-supervised learning auxiliary task, determine a unified optimization objective function, and use multi-task learning training to obtain the feature vectors of users and items;

[0020] Step 7: Target user recommendation;

[0021] Based on the obtained feature vectors of users and items, predict the score of each target user on the non-interacted items, and sort them to obtain the item recommendation list.

[0022] As a preferred embodiment, in step 1, the process of preprocessing the interaction history data is as follows:

[0023] S11: first standardize the interaction history data, and then divide the processed data set into a training data set, a validation data set, and a test data set in a ratio of 8:1:1;

[0024] S12: Pass Represents a set of n users, through Represents a set of m items. When user u interacts with item i, let r ui =1, otherwise let r ui =0, based on the above conditions, the user-item interaction matrix R is established according to the training set, R∈{0,1} n×m .

[0025] As a preferred embodiment, in step 2, the process of high-order similarity denoising is as follows:

[0026] S21: perform similarity estimation;

[0027] S211: Construct a user-item interaction bipartite graph based on the user-item interaction matrix R in,

[0028] S212: Obtain the user-item interaction bipartite graph according to formula (1) The adjacency matrix A of

[0029]

[0030] S213: Using formula (2) in the user-item interaction bipartite graph Regularization is performed on the adjacency matrix A to obtain the first-order similarity matrix between user u and item i

[0031]

[0032] Where D represents the degree matrix of A;

[0033] S214: Obtain the similarity s between the behaviors of users u and v according to formula (3) uv , according to formula (4), we can obtain the similarity s between the behaviors of items i and j ij ; Based on the similarity s between users u and v uv and the similarity s between items i and j ij Get the user similarity matrix S respectively U and item similarity matrix S I , and then obtain the second-order similarity matrix S according to formula (5);

[0034]

[0035]

[0036] The above process of obtaining the second-order similarity matrix S is equivalent to the process of obtaining the first-order similarity matrix S according to formula (6). Perform matrix operations to obtain the second-order similarity matrix S;

[0037]

[0038] S22: perform neighborhood denoising;

[0039] S221: Top-K selection: Based on the second-order similarity matrix S, select the top K similar users for each user according to formula (7) According to formula (8), select the top K similar items for each item Based on and Get the denoised second-order similarity matrix

[0040]

[0041] In the formula, represents the indicator function; represents the set of similarities between user u and its two-hop neighbors; Represents the set of similarities between item i and its two-hop neighbors;

[0042] S222: Normalization; according to formula (9) Normalize the non-zero elements in to obtain the final denoised second-order similarity matrix

[0043]

[0044] In the formula, express The degree matrix of .

[0045] As a preferred method, in step 3, the process of constructing and initializing the initial feature vectors for all users and items is as follows:

[0046] S31: Determine the number of users and items according to the user-item interaction matrix R, and construct the initial d-dimensional user feature vectors according to the one-to-one correspondence between the IDs of users and items and the feature vectors. and item feature vector Then according to formula (10), we get the initial feature matrix E (0) ;

[0047]

[0048] S32: Initial feature matrix E (0) All eigenvectors in are initialized using the Xavier method.

[0049] As a preferred method, in step 4, the process of using an alternating graph convolutional network to learn the final feature vectors of users and items and defining the recommendation loss is as follows:

[0050] S41: Alternating propagation;

[0051] S411: In the odd-numbered layers, heterogeneous neighborhood aggregation is performed based on first-order similarity, and one-hop neighbors of the same type as the target node are aggregated. The odd-numbered layer user feature vectors are obtained according to formulas (11) and (12) respectively. and item feature vector

[0052] In the formula, represents the first-order similarity between user node u and item node i, represents the first-order similarity between item node i and user node u, and Represent the one-hop neighbors of user u and item i respectively;

[0053] S412: In the even-numbered layers, homogeneous neighborhood aggregation is performed based on second-order similarity, and the two-hop neighbors with the same type as the target node are aggregated. The even-numbered layer user feature vectors are obtained according to formulas (13) and (14) respectively. and item feature vector

[0054] In the formula, represents the second-order similarity between user node u and user node v, represents the second-order similarity between item node i and item node j, and denote the two-hop neighbors of user u and item i respectively;

[0055] S413: According to formula (15), the feature matrix E of the user or item at the lth layer after alternating propagation is obtained (l) ;

[0056]

[0057] In the formula, That is, k is a natural number;

[0058] S42: multilayer polymerization;

[0059] The aggregation strategy in formula (16) is used to perform weighted superposition on the feature vectors obtained at each layer to obtain the final feature vectors e of the user. u and the final feature vector e of the item i , thus obtaining the final representation matrix E of the user or item;

[0060]

[0061] In the formula, α l Take 1 / (L+1), where L represents the total number of propagation layers.

[0062] S43: Recommendation loss construction; use the Bayesian personalized ranking loss function in formula (17) as the recommendation main task loss

[0063]

[0064] Where σ(·) represents the sigmoid function, Represents the training data set after negative sampling, that is is the prediction of user u’s interest in item i, is the prediction of user u’s interest in item j,

[0065] As a preferred embodiment, in step 5, the process of neighborhood preservation based on second-order similarity is as follows:

[0066] S51: Obtain the distance d between the user's initial feature vector and its second-layer feature vector according to formula (18) U ;

[0067]

[0068] S52: Combining formula (15) and formula (18) to obtain formula (19);

[0069]

[0070] In the formula, Where I represents the identity matrix;

[0071] S53: Obtain the distance d between the initial feature vector of the item and its second-layer feature vector according to formula (20) I ;

[0072]

[0073] In the formula,

[0074] S54: Obtain self-supervised loss according to formula (21) Essentially, the characteristics of second-order similar nodes are similar, which indirectly realizes the neighborhood preservation based on second-order similarity;

[0075]

[0076] In step 6, the process of using multi-task learning training to obtain the feature vectors of users and items is as follows:

[0077] S61: According to formula (22), the loss of the recommended main task and self-supervised loss Combined into final loss Thus, the objective function of multi-task learning is obtained;

[0078]

[0079] Where λ1 is a hyperparameter used to control the self-supervision loss The weight of the training; λ2 is a hyperparameter used to control the regularization term; Θ represents the model parameter, that is, the initial feature vector E (0) ;

[0080] S62: According to the optimization objective in formula (22), repeat steps 4 and 5 to train the model. When there is no longer a downward trend, stop training and obtain the user feature vector and item feature vector after iterative learning, that is,

[0081] In step 7, the process of sorting and obtaining the item recommendation list is as follows:

[0082] S71: Obtain a list of users who need to be recommended;

[0083] S72: Initial feature vectors of users and items obtained through training Perform the alternating graph convolution in step 4 to obtain the final feature vectors of users and items.

[0084] S73: predict the interest level of each user to be recommended in all items according to the calculation method in formula (17), that is, obtain the recommendation score;

[0085] S74: Sort by recommendation score and obtain the K items with the highest ranking to recommend to the user.

[0086] Compared with the prior art, the present invention has the following advantages:

[0087] 1. The present invention introduces high-order similarity denoising into the graph collaborative filtering recommendation method. Through high-order similarity denoising, the negative impact of noise data in the message propagation process is reduced, thereby improving the robustness of the graph collaborative filtering model. Compared with first-order similarity denoising, second-order similarity denoising comprehensively considers information such as first-order similarity, so it can implicitly achieve first-order similarity denoising.

[0088] 2. This paper proposes a new alternating graph convolutional neural network, which avoids the spread of the negative impact of noise interaction under the message propagation mechanism by performing different neighborhood aggregation operations in odd and even layers. The network can implicitly implement high-order similarity denoising and obtain more robust feature vectors of users and items.

[0089] 3. This paper proposes a new self-supervised learning framework, which fully utilizes the second-order similarity by treating the feature vectors before and after propagation based on the second-order similarity as two views and using the mean square error loss to compare the representation of the positive sample pairs. This self-supervision does not require data enhancement and negative example generation, and is only based on the existing feature vectors, which further improves the robustness and effectiveness of the model more efficiently.

[0090] In summary, compared with the prior art, the present invention obtains a more effective denoising method from the perspective of high-order similarity, and fully utilizes reliable second-order similarity in combination with self-supervised learning technology. On this basis, a robust graph collaborative filtering method is provided to obtain more accurate item recommendation results and improve the performance of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is the overall architecture diagram of the present invention;

[0092] Figure 2 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0093] The present invention will be further described below in conjunction with the accompanying drawings.

[0094] The present invention proposes a more effective denoising method from the perspective of high-order similarity to achieve robust graph collaborative filtering. Figure 1 The overall architecture diagram of the method is given. It can be seen that it consists of three core components: similarity denoising, alternating graph convolution, and self-supervised neighborhood preservation.

[0095] Specifically, Figure 2 As shown, the present invention provides a graph collaborative filtering recommendation method based on high-order similarity denoising, comprising the following steps:

[0096] Step 1: Data preprocessing;

[0097] Obtaining massive historical interaction data between users and items through data processing equipment, and preprocessing the historical interaction data using the data processing equipment to obtain a user-item interaction matrix;

[0098] Step 2: Similarity denoising;

[0099] Construct a bipartite graph of user-item interactions, perform second-order similarity estimation based on the bipartite graph, and perform neighborhood denoising on this basis to obtain the final second-order similarity;

[0100] Step 3: Initialize feature vector;

[0101] Construct and initialize initial feature vectors for all users and items;

[0102] Step 4: Recommend the construction of the main task model;

[0103] The alternating graph convolutional network is used to obtain the final feature vectors of users and items, and the recommendation loss is defined based on this;

[0104] Step 5: Self-supervised auxiliary task construction;

[0105] Define self-supervised loss to achieve neighborhood preservation based on second-order similarity;

[0106] Step 6: Multi-task learning;

[0107] Combine the recommendation main task and the self-supervised learning auxiliary task, determine a unified optimization objective function, and use multi-task learning training to obtain the feature vectors of users and items;

[0108] Step 7: Target user recommendation;

[0109] Based on the obtained feature vectors of users and items, predict the score of each target user on the non-interacted items, and sort them to obtain the item recommendation list.

[0110] As a preferred embodiment, in step 1, the process of preprocessing the interaction history data is as follows:

[0111] S11: first standardize the interaction history data, and then divide the processed data set into a training data set, a validation data set, and a test data set in a ratio of 8:1:1;

[0112] S12: Pass Represents a set of n users, through Represents a set of m items. When user u interacts with item i, let r ui =1, otherwise let r ui =0, based on the above conditions, the user-item interaction matrix R is established according to the training set, R∈{0,1} n×m .

[0113] Specifically, taking the Gowalla dataset as an example, this dataset contains the check-in behavior of users and places. A 10-core preprocessing strategy is used for this dataset, that is, users and items with less than 10 interactions are removed. After preprocessing, the number of users in the dataset is 29,859, the number of items (i.e., places) is 40,989, and the interaction data of users and items (i.e., the check-in behavior data of users and places) is 1,027,464; then, the interaction data after preprocessing is randomly divided, of which 80% is training data, 10% is test data, and the rest is verification data; based on the training data, the user-item interaction matrix R can be constructed.

[0114] As a preferred embodiment, in step 2, the process of high-order similarity denoising is as follows:

[0115] S21: perform similarity estimation;

[0116] S211: Construct a user-item interaction bipartite graph based on the user-item interaction matrix R in, It can be seen from this that Contains two types of nodes: user nodes and item nodes; if r ui =1 There is an edge between user node u and item node i, otherwise there is no edge;

[0117] S212: Obtain the user-item interaction bipartite graph according to formula (1) The adjacency matrix A of

[0118]

[0119] S213: Using formula (2) in the user-item interaction bipartite graph Regularization is performed on the adjacency matrix A to obtain the first-order similarity matrix between user u and item i

[0120]

[0121] Where D represents the degree matrix of A; therefore The rows represent the normalized interaction features of users / items;

[0122] S214: Obtain the similarity s between the behaviors of users u and v according to formula (3) uv , according to formula (4), we can obtain the similarity s between the behaviors of items i and j ij ; Based on the similarity s between users u and v uv and the similarity s between items i and j ij Get the user similarity matrix S respectively U and item similarity matrix S I , and then obtain the second-order similarity matrix S according to formula (5);

[0123]

[0124] The above process of obtaining the second-order similarity matrix S is equivalent to the process of obtaining the first-order similarity matrix S according to formula (6). Perform matrix operations to obtain the second-order similarity matrix S;

[0125]

[0126] S22: Perform neighborhood denoising; the second-order similarity S estimated above represents the similarity between the user (item) and its two-hop neighbors. Therefore, a two-stage relative denoising strategy is used to denoise the second-order similarity. The specific process is as follows:

[0127] S221: Top-K selection: Based on the second-order similarity matrix S, select the top K similar users for each user according to formula (7) According to formula (8), select the top K similar items for each item Based on and Get the denoised second-order similarity matrix

[0128]

[0129] In the formula, Represents an indicator function (i.e., returns 1 when the condition is met, otherwise 0); represents the set of similarities between user u and its two-hop neighbors; Represents the set of similarities between item i and its two-hop neighbors; it should be noted that the two-hop neighbors of the target node do not include itself;

[0130] S222: Normalization; according to formula (9) Normalize the non-zero elements in to obtain the final denoised second-order similarity matrix

[0131]

[0132] In the formula, express The degree matrix of .

[0133] As a preferred method, in step 3, the process of constructing and initializing the initial feature vectors for all users and items is as follows:

[0134] S31: Determine the number of users and items according to the user-item interaction matrix R, and construct the initial d-dimensional user feature vectors according to the one-to-one correspondence between the IDs of users and items and the feature vectors. and item feature vector Then according to formula (10), we get the initial feature matrix E (0) ;

[0135]

[0136] S32: Initial feature matrix E (0) All eigenvectors in are initialized using the Xavier method.

[0137] As a preferred method, in step 4, alternating graph convolution is the key core. Its main goal is to obtain the final feature vector based on the similarity obtained in steps 2 and 3 and the initial feature vector through the alternating graph neural network, which can be formalized as: Among them, E (0) is the initial feature matrix obtained in step 3, and They represent the regularized first-order similarity matrix obtained in step 2 and the denoised second-order similarity matrix respectively. Specifically, the process of using the alternating graph convolutional network to learn the final feature vectors of users and items and define the recommendation loss is as follows:

[0138] S41: Alternating propagation: Based on the first-order similarity and the second-order similarity, an alternating propagation strategy is used on the user-item bipartite graph to obtain the feature vector of user u and item i at the l1th or l2th layer;

[0139] S411: In the odd-numbered layers, heterogeneous neighborhood aggregation is performed based on first-order similarity, and one-hop neighbors of the same type as the target node are aggregated. The odd-numbered layer user feature vectors are obtained according to formulas (11) and (12) respectively. and item feature vector

[0140] In the formula, represents the first-order similarity between user node u and item node i, represents the first-order similarity between item node i and user node u, and Represent the one-hop neighbors of user u and item i respectively;

[0141] S412: In the even-numbered layers, homogeneous neighborhood aggregation is performed based on second-order similarity, and the two-hop neighbors with the same type as the target node are aggregated. The even-numbered layer user feature vectors are obtained according to formulas (13) and (14) respectively. and item feature vector

[0142] In the formula, represents the second-order similarity between user node u and user node v, represents the second-order similarity between item node i and item node j, and denote the two-hop neighbors of user u and item i respectively;

[0143] S413: According to formula (15), the feature matrix E of the user or item at the lth layer after alternating propagation can be obtained (l) ;

[0144]

[0145] In the formula, That is, k is a natural number;

[0146] S42: Multi-layer aggregation: After L layers of propagation, L+1 feature vectors can be obtained for each user or item. In order to obtain the final user or item representation, the following aggregation strategy is adopted:

[0147] The aggregation strategy in formula (16) is used to perform weighted superposition on the feature vectors obtained at each layer to obtain the final feature vectors e of the user. u and the final feature vector e of the item i , thus obtaining the final representation matrix E of the user or item;

[0148]

[0149] In the formula, α lTake 1 / (L+1), where L represents the total number of propagation layers.

[0150] S43: Recommendation loss construction; use the Bayesian personalized ranking loss function in formula (17) as the recommendation main task loss

[0151]

[0152] Where σ(·) represents the sigmoid function, Represents the training data set after negative sampling, that is is the prediction of user u’s interest in item i, is the prediction of user u’s interest in item j,

[0153] As a preferred method, in step five, the feature vectors before and after the first propagation based on second-order similarity are regarded as two views, and the feature similarity of the second-order similar nodes is essentially achieved by minimizing the positive example distance as the self-supervision loss, that is, the neighborhood preservation based on second-order similarity is indirectly achieved. Its main goal is to make full use of reliable second-order similarities, construct self-supervised auxiliary tasks, and improve the robustness of the model. According to step two, reliable second-order similarity can be obtained, but after steps three and four, the feature vectors obtained will weaken this similarity. In order to solve this problem, we constructed a self-supervised auxiliary task, which minimizes the distance between the initial feature vector of the user (or item) and the second-layer feature vector, so that the initial feature vectors of the user (or item) and its second-order similar users (or items) are as similar as possible. Specifically, the process of neighborhood preservation based on second-order similarity is as follows:

[0154] S51: Obtain the distance d between the user's initial feature vector and its second-layer feature vector according to formula (18) U ;

[0155]

[0156] S52: Combining formula (15) and formula (18) to obtain formula (19);

[0157]

[0158] In the formula, Where I represents the identity matrix; thus, it can be obtained that the minimization formula (18) can make the second-order similar user initial features smoother.

[0159] S53: Similarly, the distance d between the initial feature vector of the item and its second-layer feature vector can be obtained according to formula (20):I ;

[0160]

[0161] In the formula, Therefore, minimizing formula (20) can make second-order similar items smoother in the initial feature space.

[0162] S54: Obtain self-supervised loss according to formula (21) Essentially, the features of second-order similar nodes are similar, which indirectly realizes the neighborhood preservation based on second-order similarity, thereby achieving the simultaneous smoothing of second-order similar user initial features and item initial features;

[0163]

[0164] As a preferred embodiment, in step six, a multi-task learning strategy is adopted to combine the recommendation main task and the self-supervised auxiliary task. Specifically, the process of obtaining the feature vectors of users and items by multi-task learning training is as follows:

[0165] S61: According to formula (22), the loss of the recommended main task and self-supervised loss Combined into final loss Thus, the objective function of multi-task learning is obtained;

[0166]

[0167] Where λ1 is a hyperparameter used to control the self-supervision loss The weight of the training; λ2 is a hyperparameter used to control the regularization term; Θ represents the model parameter, that is, the initial feature vector E (0) ;

[0168] S62: According to the optimization objective in formula (22), repeat steps 4 and 5 to train the model. When there is no longer a downward trend, stop training and obtain the user feature vector and item feature vector after iterative learning, that is,

[0169] As a preference, in one embodiment of the present invention, the maximum number of iteration rounds selected is 300. The training is stopped when there is no longer a downward trend within 10 rounds, so as to obtain the final user and item initial feature vectors for the final recommendation effect test.

[0170] As a preferred embodiment, in step seven, the process of sorting and obtaining the item recommendation list is as follows:

[0171] S71: Obtain a list of users who need to be recommended;

[0172] S72: Initial feature vectors of users and items obtained through training Perform the alternating graph convolution in step 4 to obtain the final feature vectors of users and items.

[0173] S73: predict the interest level of each user to be recommended in all items according to the calculation method in formula (17), that is, obtain the recommendation score;

[0174] S74: Sort by recommendation score and obtain the K items with the highest ranking to recommend to the user.

[0175] The present invention provides a collaborative filtering recommendation method based on high-order similarity denoising, which introduces high-order similarity denoising into the graph collaborative filtering recommendation method, performs denoising based on reliable second-order similarity, reduces the negative impact of noise data in the message propagation process, and thus improves the robustness of the graph collaborative filtering model. Compared with the first-order similarity denoising representing the interaction between users and items, the second-order similarity denoising comprehensively considers information such as the first-order similarity, so the first-order similarity denoising can be implicitly realized. In addition, on the basis of similarity denoising, a new alternating graph convolutional neural network is proposed, which avoids the spread of the negative impact of noise interaction under the message propagation mechanism by performing different neighborhood aggregation operations in odd and even layers, and obtains more robust user and item feature vectors. At the same time, the present invention also proposes a new self-supervised learning framework, which regards the feature vectors before and after propagation based on the second-order similarity as two views, uses the mean square error loss to compare the representation of the positive sample pair, and realizes the full utilization of the second-order similarity. The self-supervision does not require data enhancement and negative example generation, and is only based on the existing feature vectors, which further improves the robustness and effectiveness of the model more efficiently.

[0176] In summary, compared with the prior art, the present invention obtains a more effective denoising method from the perspective of high-order similarity, and fully utilizes reliable second-order similarity in combination with self-supervised learning technology. On this basis, a robust graph collaborative filtering method is provided to obtain more accurate item recommendation results and improve the performance of the recommendation system.

Claims

1. A graph collaborative filtering recommendation method based on high-order similarity denoising, characterized in that: The following steps are involved: Step 1: Data preprocessing; Obtaining massive historical interaction data between users and items through data processing equipment, and preprocessing the historical interaction data using the data processing equipment to obtain a user-item interaction matrix; Step 2: Similarity denoising; Construct a bipartite graph of user-item interactions, perform second-order similarity estimation based on the bipartite graph, and perform neighborhood denoising on this basis to obtain the final second-order similarity; Step 3 : Feature vector initialization; Construct and initialize initial feature vectors for all users and items; Step 4: Recommend the construction of the main task model; The alternating graph convolutional network is used to obtain the final feature vectors of users and items, and the recommendation loss is defined based on this; Step 5: Self-supervised auxiliary task construction; Define self-supervised loss to achieve neighborhood preservation based on second-order similarity; Step 6: Multi-task learning; Combine the recommendation main task and the self-supervised learning auxiliary task, determine a unified optimization objective function, and use multi-task learning training to obtain the feature vectors of users and items; Step 7: Target user recommendation; Based on the obtained feature vectors of users and items, predict the score of each target user on the non-interacted items, and sort them to obtain the item recommendation list.

2. According to claim 1, a graph collaborative filtering recommendation method based on high-order similarity denoising is characterized in that: In step 1, the process of preprocessing the interaction history data is as follows: S11: first standardize the interaction history data, and then divide the processed data set into a training data set, a validation data set, and a test data set in a ratio of 8:1:1; S12: Pass Represents a set of n users, through Represents a set of m items. When user u interacts with item i, let r ui =1, otherwise let r ui =0, based on the above conditions, the user-item interaction matrix R is established according to the training set, R∈{0,1} n×m .

3. The graph collaborative filtering recommendation method based on high-order similarity denoising according to claim 2, characterized in that: In step 2, the process of high-order similarity denoising is as follows: S21: perform similarity estimation; S211: Construct a user-item interaction bipartite graph based on the user-item interaction matrix R in, S212: Obtain the user-item interaction bipartite graph according to formula (1) The adjacency matrix A of S213: Using formula (2) in the user-item interaction bipartite graph Regularization is performed on the adjacency matrix A to obtain the first-order similarity matrix between user u and item i Where D represents the degree matrix of A; S214: Obtain the similarity s between the behaviors of users u and v according to formula (3) uv , according to formula (4), we can obtain the similarity s between the behaviors of items i and j ij ; Based on the similarity s between users u and v uv and the similarity s between items i and j ij Get the user similarity matrix S respectively U and item similarity matrix S I , and then obtain the second-order similarity matrix S according to formula (5); The above process of obtaining the second-order similarity matrix S is equivalent to the process of obtaining the first-order similarity matrix S according to formula (6). Perform matrix operations; S22: perform neighborhood denoising; S221: Top-K selection: Based on the second-order similarity matrix S, select the top K similar users for each user according to formula (7) According to formula (8), select the top K similar items for each item Based on and Get the denoised second-order similarity matrix In the formula, represents the indicator function; represents the set of similarities between user u and its two-hop neighbors; represents the set of similarities between item i and its two-hop neighbors; S222: Normalization; according to formula (9) Normalize the non-zero elements in to obtain the final denoised second-order similarity matrix In the formula, express The degree matrix of .

4. The graph collaborative filtering recommendation method based on high-order similarity denoising according to claim 3 is characterized in that: In step 3, the process of constructing and initializing the initial feature vectors for all users and items is as follows: S31: Determine the number of users and items according to the user-item interaction matrix R, and construct the initial d-dimensional user feature vectors according to the one-to-one correspondence between the IDs of users and items and the feature vectors. and item feature vector Then according to formula (10), we get the initial feature matrix E (0) ; S32: Initial feature matrix E (0) All eigenvectors in are initialized using the Xavier method.

5. According to claim 4, a graph collaborative filtering recommendation method based on high-order similarity denoising is characterized in that: In step 4, the process of using the alternating graph convolutional network to obtain the final feature vectors of users and items and define the recommendation loss is as follows: S41: Alternating propagation; S411: In the odd-numbered layers, heterogeneous neighborhood aggregation is performed based on first-order similarity, and one-hop neighbors of the same type as the target node are aggregated. The odd-numbered layer user feature vectors are obtained according to formulas (11) and (12) respectively. and item feature vector In the formula, represents the first-order similarity between user node u and item node i, represents the first-order similarity between item node i and user node u, and denote the one-hop neighbors of user u and item i respectively; S412: In the even-numbered layers, homogeneous neighborhood aggregation is performed based on second-order similarity, and the two-hop neighbors with the same type as the target node are aggregated. The even-numbered layer user feature vectors are obtained according to formulas (13) and (14) respectively. and item feature vector In the formula, represents the second-order similarity between user node u and user node v, represents the second-order similarity between item node i and item node j, and Represent the two-hop neighbors of user u and item i respectively; S413: According to formula (15), the feature matrix E of the user or item at the lth layer after alternating propagation is obtained (l) ; In the formula, That is, k is a natural number; S42: multilayer polymerization; The aggregation strategy in formula (16) is used to perform weighted superposition on the feature vectors obtained at each layer to obtain the final feature vectors e of the user. u and the final feature vector e of the item i , thus obtaining the final representation matrix E of the user or item; In the formula, α l Take 1 / (L+1), where L represents the total number of propagation layers. S43: Recommendation loss construction; use the Bayesian personalized ranking loss function in formula (17) as the recommendation main task loss Where σ(·) represents the sigmoid function, Represents the training data set after negative sampling, that is is the prediction of user u’s interest in item i, is the prediction of user u’s interest in item j, 6. The graph collaborative filtering recommendation method based on high-order similarity denoising according to claim 5, characterized in that: In step 5, the process of neighborhood preservation based on second-order similarity is as follows: S51: Obtain the distance d between the user's initial feature vector and its second-layer feature vector according to formula (18) U ; S52: Combining formula (15) and formula (18) to obtain formula (19); In the formula, Where I represents the identity matrix; S53: Obtain the distance d between the initial feature vector of the item and its second-layer feature vector according to formula (20) O ; In the formula, S54: Obtain self-supervised loss according to formula (21) Essentially, the characteristics of second-order similar nodes are similar, which indirectly realizes the neighborhood preservation based on second-order similarity; 7. The graph collaborative filtering recommendation method based on high-order similarity denoising according to claim 6, characterized in that: In step 6, the process of using multi-task learning training to obtain the feature vectors of users and items is as follows: S61: According to formula (22), the loss of the recommended main task and self-supervised loss Combined into final loss Thus, the objective function of multi-task learning is obtained; Where λ1 is a hyperparameter used to control the self-supervision loss The weight of the training; λ2 is a hyperparameter used to control the regularization term; Θ represents the model parameter, that is, the initial feature vector E (0) ; S62: According to the optimization objective in formula (22), repeat steps 4 and 5 to train the model. When there is no longer a downward trend, stop training and obtain the user feature vector and item feature vector after iterative learning, that is, 8. The graph collaborative filtering recommendation method based on high-order similarity denoising according to claim 7, characterized in that: In step 7, the process of sorting and obtaining the item recommendation list is as follows: S71: Obtain a list of users who need to be recommended; S72: Initial feature vectors of users and items obtained through training Perform the alternating graph convolution in step 4 to obtain the final feature vectors of users and items. S73: predict the interest level of each user to be recommended in all items according to the calculation method in formula (17), that is, obtain the recommendation score; S74: Sort by recommendation score and obtain the K items with the highest ranking to recommend to the user.

Images