Graph denoising collaborative filtering recommendation method based on alignment and uniformity

Abstract

The invention discloses a graph denoising collaborative filtering recommendation method based on alignment and uniformity. The method mainly comprises the following steps: firstly, on a user-article interaction graph, pre-training a LightGCN model to obtain user / article embedding characterization; secondly, according to a pre-trained user / article embedding inner product, selecting a high-similarity user / article to reconstruct a user-article interaction graph; then, based on the reconstructed user-article interaction diagram, using LightGCN to learn user / article characterization containing a high-order cooperative signal; then, calculating alignment loss and uniform loss, constructing an objective function of a graph denoising collaborative filtering model, and learning model parameters by using an Adam optimizer; and finally, using the inner product of the learned user / article embedded representation as a prediction score, and performing personalized recommendation according to the prediction score. According to the method, a graph denoising module based on alignment and uniformity is integrated in a traditional graph collaborative filtering algorithm, the module not only utilizes alignment and uniformity operation to relieve the problem of feature collapse, but also adopts graph denoising operation to eliminate noise interference in an original interaction graph, and recommendation performance is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence recommendation systems, and in particular to a graph denoising collaborative filtering recommendation method based on alignment and uniformity. Background Art

[0002] Recommendation algorithms based on collaborative filtering exploit users' historical behavior data to identify potential preferences and recommend items they may be interested in. Classic collaborative filtering-based recommendation algorithms include MF, NCF, and LRML. However, these methods lack the ability to encode high-order collaborative signals, resulting in recommendation models that fail to fully utilize the interaction between users and items, leading to suboptimal recommendation performance.

[0003] In recent years, graph convolutional neural networks (GCNNs) have captured high-order collaborative signals in user-item interaction graphs and learned the underlying features of entities, thereby providing recommendation models with rich user and item representations and effectively improving recommendation quality. Specifically, GCNNs learn the representation of the current node by aggregating the representations of neighboring nodes, making the representations of the node as similar as possible to those of its neighbors. However, this operation leads to high correlation between node representations in the user-item interaction graph, which can easily lead to feature collapse.

[0004] Alignment and uniform loss are potential solutions to mitigate feature collapse. Alignment loss minimizes the distance between positive pairs, placing users and items they interact with as close together as possible in feature space. Uniform loss maximizes the distance between pairs in feature space, ensuring that all user and item representations are evenly distributed on the unit hypersphere, alleviating the problem of feature collapse to some extent.

[0005] This method integrates a graph denoising module based on alignment and uniformity into traditional graph collaborative filtering algorithms. This module not only uses alignment and uniformity operations to alleviate the problem of feature collapse, but also employs graph denoising to remove noise data from the original interaction graph, effectively improving the performance of existing recommendation algorithms. Summary of the Invention

[0006] The technical problem addressed by this invention is that traditional graph collaborative filtering suffers from the interference of noisy data and feature collapse in the feature space. As a result, traditional graph collaborative filtering algorithms are unable to fully learn the embedded representations of users and items, resulting in suboptimal recommendation performance. This invention integrates a graph denoising module based on alignment and uniformity into the traditional graph collaborative filtering recommendation algorithm, thereby improving the performance of the recommendation algorithm.

[0007] The present invention adopts the following technical solutions to solve the above technical problems:

[0008] 1. A graph denoising collaborative filtering recommendation method based on alignment and uniformity, characterized by comprising the following steps:

[0009] Step 1) Pre-train the LightGCN model on the user-item interaction graph to obtain user / item embedding representations;

[0010] Step 2) Select high-similarity users / items based on the pre-trained user / item embedding inner product to reconstruct the user-item interaction graph;

[0011] Step 3) Based on the reconstructed user-item interaction graph, LightGCN is used to learn user / item representations containing high-order collaborative signals;

[0012] Step 4) Calculate the alignment loss and uniform loss, construct the objective function of the graph denoising collaborative filtering model, and use the Adam optimizer to learn the model parameters;

[0013] Step 5) Use the inner product of the learned user / item embedding representation as the predicted score and make personalized recommendations based on the predicted score.

[0014] 2. The alignment and uniformity-based graph denoising collaborative filtering recommendation method according to claim 1 is characterized in that, in step 1), a pre-trained LightGCN model is used on the user-item interaction graph to obtain user / item embedding representations.

[0015] First, the implicit user feature matrix and the implicit item feature matrix are initialized by Gaussian prior as and d represents the embedding dimension, M and N represent the number of users and items respectively, as follows:

[0016] U~N(0,0.01)

[0017] I~N(0,0.01)

[0018] Among them, U and I obey the normal distribution with mean 0 and variance 0.01.

[0019] Given the interaction information between users and items, the embedding representations of user u and item i are obtained through the feature matrices U and I:

[0020] e u =lookup(U,u)

[0021] e i =lookup(I,i)

[0022] in, u and i represent the user and item indexes, respectively. The lookup(·) operation extracts the corresponding embedding vector from the corresponding embedding representation matrix based on the index information of the entity.

[0023] Based on the user and item interaction information, construct the user and item interaction matrix If user u has interacted with item i, the value of row u and column i in the interaction matrix R between user u and item i is 1, otherwise it is 0. The corresponding adjacency matrix A is constructed using the interaction matrix R between user and item as follows:

[0024]

[0025] Among them, R T is the transposed matrix of R.

[0026] Next, based on the adjacency matrix A, construct the Laplace matrix L as follows:

[0027]

[0028] Where D is a diagonal matrix, formalized as:

[0029]

[0030] in, represents the number of first-order neighbors of user u1, Indicates the number of first-order neighbors of item i1.

[0031] After the graph convolution operation, the embedding representation of users and items at layer l is as follows:

[0032] E (l) =LE (l-1)

[0033] in is the initial embedding representation, represents the initial embedding representation of user u1, Represents the initial embedding representation of item i1.

[0034] After L layers of convolution, we obtain the final embedding representation as follows:

[0035] E * =E (0) +E (1) +…+E (L)

[0036] E * The final user embedding representation U is obtained by dividing the number of users M and the number of items N. * And the final item embedding representation I * .

[0037] Then, according to U * and I * , the final embedding representation of user u and item i is extracted through the kookup(·) operation as follows:

[0038]

[0039] In addition, by calculating and The inner product of gets the predicted score y of user u for item i u,i ,as follows:

[0040]

[0041] Finally, we minimize the objective function based on the BPR loss To learn the model parameters, as follows:

[0042]

[0043] Where σ(x)=1 / (1+e -x ). D s Represents the training data set, where each sample is a triple (u,i + ,i - ), i + Indicates the items that user u has interacted with, i - represents the items that user u has not interacted with, y u,i+ represents the score of the user-item positive sample pair, y u,i- represents the score of the user-item negative sample pair, λ represents the regularization coefficient, and φ={U,I} is the parameter of the model. In addition, we use the Adam optimizer to minimize And save the embedded representations of users and items learned after training.

[0044] 3. The method according to claim 2 is characterized in that in step 2), users / items with high similarity are selected based on the inner product of pre-trained user / item embeddings to reconstruct the user-item interaction graph.

[0045] First, read the embedded representation of user u and item i obtained by pre-training and Use the inner product to calculate the similarity score between user u and item, and item i and user, and extract the corresponding top-K users or items according to the similarity score, as follows:

[0046]

[0047] in, and Represent the user set and item set respectively, e u and e i are the embedding representations of user u and item i read from the pre-trained model, and represents the final user and item embedding matrix read from the pre-trained model, U K and I K They represent the hyperparameters controlling the number of items selected by users and the number of users selecting items, respectively.

[0048] Reconstruct the user-item interaction matrix based on the extracted top-K user or item information If user u has interacted with item i, the interaction matrix between user u and item i is The value of row u and column i in is 1, otherwise it is 0.

[0049] 4. The graph denoising collaborative filtering recommendation method based on alignment and uniformity according to claim 3 is characterized in that in step 3), based on the reconstructed user-item interaction graph, LightGCN is used to learn user / item representations containing high-order collaborative signals.

[0050] First, using the user-item interaction matrix Construct the corresponding adjacency matrix as follows:

[0051]

[0052] in, for The transposed matrix of .

[0053] According to the reconstructed adjacency matrix Construct a new Laplacian matrix as follows:

[0054]

[0055] in, is the diagonal matrix after denoising, which is formalized as:

[0056]

[0057] in, represents the number of first-order neighbors of user u1 after denoising, Indicates the number of first-order neighbors of item i1 after denoising.

[0058] After the graph convolution operation, the embedding representation of users and items at layer l is as follows:

[0059]

[0060] in is the initial embedding representation, represents the initial embedding representation of user u1 after denoising, Represents the initial embedding representation of item i1 after denoising.

[0061] After L layers of convolution, we will obtain the final embedding representation as follows:

[0062]

[0063] Will Split according to the number of users M and the number of items N to obtain the final denoised user embedding representation And the final embedding representation of all denoised items

[0064] Finally, according to and Get the final embedding representation of user u and item i as follows:

[0065]

[0066] 5. The graph denoising collaborative filtering recommendation method based on alignment and uniformity according to claim 4, characterized in that in step 4), the alignment loss and uniformity loss are calculated, the objective function of the graph denoising collaborative filtering model is constructed, and the model parameters are learned using the Adam optimizer.

[0067] First, the alignment loss is calculated, which is formalized as follows:

[0068]

[0069] Among them, (u,i) represents the tuple of user u interacting with item i, p pos represents the distribution of positive sample pairs, f(·) is the L2 norm, ||·|| 2 represents the Euclidean distance.

[0070] Next, we calculate the uniformity loss, which is formalized as follows:

[0071]

[0072] Among them, p user represents the distribution of user sample pairs, p item represents the distribution of item sample pairs, u and u' represent the user distribution p user Two different users in, i and i' respectively represent the item distribution p item Two different items in .

[0073] Finally, by optimizing the alignment loss and uniformity loss simultaneously, the objective formula of the graph denoising collaborative filtering recommendation algorithm based on alignment and uniformity is obtained as follows:

[0074]

[0075] Where γ is the weight coefficient for adjusting the uniform loss, which is minimized by the Adam optimizer

[0076] 6. The alignment and uniformity-based graph denoising collaborative filtering recommendation method according to claim 5, wherein in step 5), the inner product of the learned user / item embedding representation is used as the predicted score, and personalized recommendations are performed based on the predicted score.

[0077] Based on the embedding representation of users and items and Get user u’s preference prediction score y for item i u,i ,as follows:

[0078]

[0079] Where ⊙ represents the vector inner product.

[0080] Finally, after calculating the predicted scores for all users and items, the top items are recommended to users according to the score ranking. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 The figure shows a flowchart of a graph denoising collaborative filtering recommendation method based on alignment and uniformity.

Claims

1. A graph denoising collaborative filtering recommendation method based on alignment and uniformity, characterized by: The following steps are involved: Step 1) Pre-train the LightGCN model on the user-item interaction graph to obtain user / item embedding representations; Step 2) Select high-similarity users / items based on the pre-trained user / item embedding inner product to reconstruct the user-item interaction graph; Step 3) Based on the reconstructed user-item interaction graph, LightGCN is used to learn user / item representations containing high-order collaborative signals; Step 4) Calculate the alignment loss and uniform loss, construct the objective function of the graph denoising collaborative filtering model, and use the Adam optimizer to learn the model parameters; Step 5) Use the inner product of the learned user / item embedding representation as the predicted score and make personalized recommendations based on the predicted score.

2. The alignment and uniformity-based graph denoising collaborative filtering recommendation method according to claim 1 is characterized in that, in step 1), a pre-trained LightGCN model is used on the user-item interaction graph to obtain user / item embedding representations. First, the implicit user feature matrix and the implicit item feature matrix are initialized by Gaussian prior as and d represents the embedding dimension, M and N represent the number of users and items respectively, as follows: U~N(0,0.01) I~N(0,0.01) in, U and I follow a normal distribution with a mean of 0 and a variance of 0.

01. Given the interaction information between users and items, the embedding representations of user u and item i are obtained through the feature matrices U and I: e u =lookup(U,u) e i =lookup(I,i) in, u and i represent the user and item indexes, respectively. The lookup(·) operation extracts the corresponding embedding vector from the corresponding embedding representation matrix based on the index information of the entity. Based on the user and item interaction information, construct the user and item interaction matrix If user u has interacted with item i, the value of row u and column i in the interaction matrix R between user u and item i is 1, otherwise it is 0. The corresponding adjacency matrix A is constructed using the interaction matrix R between users and items as follows: Among them, R T is the transposed matrix of R. Next, based on the adjacency matrix A, construct the Laplace matrix L as follows: Where D is a diagonal matrix, formalized as: in, represents the number of first-order neighbors of user u1, Indicates the number of first-order neighbors of item i1. After the graph convolution operation, the embedding representation of users and items at layer l is as follows: AND (l) =THE (l-1) in is the initial embedding representation, represents the initial embedding representation of user u1, Represents the initial embedding representation of item i1. After L layers of convolution, we obtain the final embedding representation as follows: AND * =And (0) +E (1) +…+And (L) E * The final user embedding representation U is obtained by dividing the number of users M and the number of items N. * And the final item embedding representation I * . Then, according to U * and I * , the final embedding representation of user u and item i is extracted through the lookuo(·) operation as follows: In addition, by calculating and The inner product of gets the predicted score y of user u for item i u,i ,as follows: Finally, we minimize the objective function based on the BPR loss To learn the model parameters, as follows: Where σ(x)=1 / (1+e -x ). D s Represents the training data set, where each sample is a triple (u,i + ,i - ), i + Indicates the items that user u has interacted with, i - represents items that user u has not interacted with, represents the score of the user-item positive sample pair, represents the score of the user-item negative sample pair, λ represents the regularization coefficient, and φ={U,I} is the parameter of the model. In addition, we use the Adam optimizer to minimize And save the embedded representations of users and items learned after training.

3. The method according to claim 2 is characterized in that in step 2), users / items with high similarity are selected based on the pre-trained user / item embedding inner product to reconstruct the user-item interaction graph. First, read the embedded representation of user u and item i obtained by pre-training and Use the inner product to calculate the similarity score between user u and item, and item i and user, and extract the corresponding top-K users or items according to the similarity score, as follows: in, and Represent the user set and item set respectively, e u and e i are the embedding representations of user u and item i read from the pre-trained model, and represents the final user and item embedding matrix read from the pre-trained model, U K and I K They represent the hyperparameters controlling the number of items selected by users and the number of users selecting items, respectively. Reconstruct the user-item interaction matrix based on the extracted top-K user or item information If user u has interacted with item i, the interaction matrix between user u and item i is The value of row u and column i in is 1, otherwise it is 0.

4. The graph denoising collaborative filtering recommendation method based on alignment and uniformity according to claim 3 is characterized in that: In step 3), based on the reconstructed user-item interaction graph, LightGCN is used to learn user / item representations containing high-order collaborative signals. First, using the user-item interaction matrix Construct the corresponding adjacency matrix as follows: in, for The transposed matrix of . According to the reconstructed adjacency matrix Construct a new Laplacian matrix as follows: in, is the diagonal matrix after denoising, which is formalized as: in, represents the number of first-order neighbors of user u1 after denoising, Indicates the number of first-order neighbors of item i1 after denoising. After the graph convolution operation, the embedding representation of users and items at layer l is as follows: in is the initial embedding representation, represents the initial embedding representation of user u1 after denoising, Represents the initial embedding representation of item i1 after denoising. After L layers of convolution, we will obtain the final embedding representation as follows: Will Split according to the number of users M and the number of items N to obtain the final denoised user embedding representation And the final embedding representation of all denoised items Finally, according to and Get the final embedding representation of user u and item i as follows:

5. The graph denoising collaborative filtering recommendation method based on alignment and uniformity according to claim 4, characterized in that: In step 4), the alignment loss and uniform loss are calculated, the objective function of the graph denoising collaborative filtering model is constructed, and the model parameters are learned using the Adam optimizer. First, the alignment loss is calculated, which is formalized as follows: Among them, (u,i) represents the tuple of user u interacting with item i, p pos represents the distribution of positive sample pairs, f(·) is the L2 norm, ||·|| 2 represents the Euclidean distance. Next, we calculate the uniformity loss, which is formalized as follows: Among them, p user represents the distribution of user sample pairs, p item represents the distribution of item sample pairs, u and u' represent the user distribution p user Two different users in, i and i' respectively represent the item distribution p item Two different items in . Finally, by optimizing the alignment loss and uniformity loss simultaneously, the objective formula of the graph denoising collaborative filtering recommendation algorithm based on alignment and uniformity is obtained as follows: Where γ is the weight coefficient for adjusting the uniform loss, which is minimized by the Adam optimizer 6. The graph denoising collaborative filtering recommendation method based on alignment and uniformity according to claim 5, characterized in that: In step 5), the inner product of the learned user / item embedding representation is used as the predicted score, and personalized recommendations are performed based on the predicted score. Based on the embedding representation of users and items and Get user u’s preference prediction score y for item i u,i ,as follows: Where ⊙ represents the vector inner product. Finally, after calculating the predicted scores for all users and items, the top items are recommended to users according to the score ranking.

Images