Popularity bias elimination recommendation method based on odd-even layer graph contrast learning and cross-pair sampling

Abstract

The present invention discloses a recommendation method for eliminating popularity bias based on odd-even layer graph contrast learning and cross-pair sampling. The method comprises: constructing a bipartite graph based on an acquired dataset of user-item interactions; inputting the adjacency matrix of the bipartite graph into a graph convolutional network in a debiased recommendation model, and taking the output mean of all convolutional layers as the original views of the user and item; inputting the adjacency matrix of the bipartite graph into a noise-perturbed graph convolutional network in the debiased recommendation model, and constructing contrast views based on the neighbor structure relationship of nodes in odd and even layers of the convolution; cross-pair sampling is used on the dataset to obtain a training set; calculating the loss function of the main task based on the training set and the original views, and calculating the loss function of the proxy task based on the contrast views; jointly optimizing the two loss functions to obtain a final node representation; and obtaining an unbiased item recommendation list corresponding to the user based on the final node representation. The present invention can effectively eliminate popularity bias in recommendation results and effectively improve the diversity of recommendation results while ensuring recommendation accuracy.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a popularity bias elimination recommendation method, system, electronic device and storage medium based on odd-even layer graph comparative learning and cross-pairing sampling. Background Art

[0002] With the rapid development of science and technology, the amount of information on the internet is growing rapidly, and information overload is becoming increasingly serious. Leveraging this massive amount of data to provide personalized recommendations has become a pressing challenge for the industry. As a tool for discovering users' potential interests and streamlining recommendation strategies, recommender systems have been widely adopted by major e-commerce platforms. They predict users' potential products, provide them with a positive shopping experience, and generate incremental revenue and user engagement for e-commerce platforms.

[0003] However, due to the inherent characteristics of recommendation systems, biases can arise in every step of the recommendation process. These biases can exacerbate over time in the feedback loop, thus affecting the accuracy of recommendations. Therefore, eliminating bias in recommendation systems and improving recommendation accuracy has gradually become a popular research topic. Eliminating popularity bias is a major area of ​​research in recommendation system bias. Popularity bias is defined as the frequency with which popular items are recommended exceeding their true popularity. This manifests as homogeneous popular items being preferentially recommended to multiple users, some of which may not match individual preferences. As can be seen, popularity bias not only compromises the diversity of recommendation results but also further increases the exposure opportunities for popular items, making interaction data more long-tailed. Summary of the Invention

[0004] In order to address the deficiencies of the above-mentioned prior art, the present invention provides a popularity bias elimination recommendation method, system, electronic device and storage medium based on odd-even layer graph contrast learning and cross-pairing sampling. During the training process, the representation distribution of nodes is effectively adjusted through the homogeneous neighbor comparison view and heterogeneous neighbor comparison view in the proxy task, so that the debiased recommendation model learns evenly distributed user and item representation vectors, thereby providing a recommendation list that better meets user needs; at the same time, cross-pairing sampling is innovatively adopted to enable the main task to achieve debiasing at the level of training samples, ensuring the excellent performance of the debiased recommendation model in terms of recommendation accuracy, debiasing ability and training efficiency; the proxy task achieves debiasing at the representation level based on the main task, and through the organic combination of the two debiasing levels, promotes the debiased recommendation model to better eliminate the popularity bias in the recommendation results, and effectively improves the diversity of the recommendation results while ensuring the accuracy of the recommendation.

[0005] The first object of the present invention is to provide a popularity bias elimination recommendation method based on odd-even layer graph contrast learning and cross-pairing sampling.

[0006] The second object of the present invention is to provide a popularity bias elimination recommendation system based on odd-even layer graph contrast learning and cross-pairing sampling.

[0007] A third object of the present invention is to provide an electronic device.

[0008] A fourth object of the present invention is to provide a storage medium.

[0009] The first object of the present invention can be achieved by adopting the following technical solutions:

[0010] A popularity bias elimination recommendation method based on odd-even layer graph contrast learning and cross-pair sampling, the method comprising:

[0011] Construct a bipartite graph based on the obtained dataset of user-item interactions;

[0012] The adjacency matrix and initial eigenvector matrix of the bipartite graph are input into the graph convolutional network in the debiased recommendation model, and the output mean of all convolutional layers is used as the original view of users and items;

[0013] The adjacency matrix of the bipartite graph is input into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer. Based on the enhanced user and item views, heterogeneous and homogeneous neighbor comparison views are constructed based on the neighbor structure relationships of nodes in odd and even convolutional layers.

[0014] Based on the dataset, a training set is obtained by cross-pair sampling; the loss function of the main task is calculated based on the training set and the original view; the loss function of the proxy task is calculated based on the heterogeneous and homogeneous neighbor comparison views; and the final node representation is obtained by jointly optimizing the loss functions of the main task and the proxy task.

[0015] The relevance scores of users and items are predicted based on the final node representation, and an unbiased item recommendation list is filtered out for the user based on the relevance scores.

[0016] Furthermore, the number of noise perturbation convolutional layers in the noise perturbation graph convolutional network is 3;

[0017] According to the enhanced user and item views, based on the neighbor structure relationship of nodes in the odd and even convolution layers, heterogeneous and homogeneous neighbor comparison views are constructed respectively, including:

[0018] The neighboring nodes of a node in an odd-numbered layer are regarded as heterogeneous neighbors, and the output of the noise-perturbed convolutional layer where the odd number is located is selected to construct a heterogeneous neighbor comparison view;

[0019] The neighboring nodes of a node in the even-numbered layer are regarded as isomorphic neighbors, and the output of the noise-perturbed convolutional layer where the even number is located is selected to construct a comparative view of isomorphic neighbors.

[0020] Furthermore, the neighboring nodes of the node in the odd-numbered layer are regarded as heterogeneous neighbors, and the output of the noise perturbation convolution layer where the odd number is located is selected to construct a heterogeneous neighbor comparison view, including:

[0021]

[0022]

[0023] Where, and The heterogeneous neighborhood comparison views of users and items respectively; V u and V i are the user node set and the item node set respectively; τ is the temperature coefficient hyperparameter; and are the normalized representations of user nodes and item nodes respectively, and are the enhanced user view and item view output by the mth noise perturbation convolutional layer, m = 0, 1, 2, 3;

[0024] The neighboring nodes of the node in the even-numbered layer are regarded as isomorphic neighbors, and the output of the noise perturbation convolution layer where the even-numbered nodes are located is selected to construct a isomorphic neighbor comparison view, including:

[0025]

[0026] Where, and Isomorphic neighborhood comparison views for users and items, respectively.

[0027] Furthermore, the adjacency matrix of the bipartite graph is input into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer, including:

[0028] The adjacency matrix and initial eigenvector matrix of the bipartite graph are input into the noise-perturbed graph convolutional network. The graph convolution operation is first performed using three noise-perturbed convolutional layers. The formula is as follows:

[0029]

[0030] Where l = 0, 1, 2; are the feature vectors of users and items output by the l+1th convolutional layer respectively; is the neighbor node set of user node u, is the set of neighbor nodes of item node i; symmetric normalization term Follow the design of standard GCN; are the feature vectors of users and items output by the lth convolutional layer respectively; and The initial feature vectors of users and items in the initial feature vector matrix are used as the output of the 0th convolutional layer;

[0031] After noise disturbance, the formula is as follows:

[0032]

[0033] Where m = 0, 1, 2, 3; and are the enhanced user and item views obtained by perturbation of the m-th convolutional layer; and is the noise applied to the user and item nodes in the mth convolutional layer; f norm (·) normalizes the random noise vector ξ to a unit vector; sign(·) restricts the direction of the normalized noise vector to (or ) in the same hyperspace quadrant; the constant parameter δ is used to limit the length of the noise.

[0034] Furthermore, the final node representation is obtained by jointly optimizing the loss functions of the main task and the proxy task, including:

[0035] According to the loss function of the main task and the loss function of the proxy task, the optimization objective is as follows:

[0036]

[0037] Where, The loss function of the main task; is the loss function of the proxy task; λ1 and λ2 are regularization parameters; Θ represents the model parameters, and ‖·‖2 is the L2 regularization function;

[0038] During the training process, the loss back propagation and gradient descent methods are used to optimize That is, the final node representation is obtained.

[0039] Furthermore, the interactions between all users and items in the dataset are cross-sampled to obtain the training set Where k is the number of interactions contained in each sample; b is the training batch; u s and i s are user nodes and product nodes respectively, s=1,2,...,bk; different samples are combined by taking multiple k;

[0040] The formula of the loss function of the main task is as follows:

[0041]

[0042] Where σ(·) is the sigmoid function; is the correlation score between the user and the item calculated by the vector inner product; e u and e i are the original views of users and items respectively.

[0043] Furthermore, the formula of the loss function of the proxy task is as follows:

[0044]

[0045] Where, is the homogeneous neighbor contrast loss function calculated based on InfoNCE and heterogeneous neighbor contrast views, is the heterogeneous neighbor contrast loss function calculated based on InfoNCE and homogeneous neighbor contrast views; the hyperparameter α is used to balance the weights of the heterogeneous neighbor contrast loss function and the homogeneous neighbor contrast loss function.

[0046] Furthermore, constructing a bipartite graph based on the acquired dataset of user-item interactions includes:

[0047] The set of user and item interactions contained in the dataset is denoted as Each element l t =(u,i) is a historical interaction between user u and item i;

[0048] Consider users and items as two types of nodes in a bipartite graph, and an interaction as an edge between two nodes to construct a bipartite graph. Node Collection Contains user node set V u and the commodity node set V i ; Edge set It is the connection relationship between user nodes and product nodes.

[0049] The second object of the present invention can be achieved by adopting the following technical solutions:

[0050] A popularity bias elimination recommendation system based on odd-even layer graph contrast learning and cross-pair sampling, the system comprising:

[0051] A construction module is used to construct a bipartite graph based on the obtained dataset of user-item interactions;

[0052] The original view generation module is used to input the adjacency matrix of the bipartite graph into the graph convolutional network in the debiased recommendation model and use the mean output of all convolutional layers as the original view of users and items;

[0053] The comparative view generation module is used to input the adjacency matrix of the bipartite graph into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer. Based on the enhanced user and item views, heterogeneous and homogeneous neighbor comparative views are constructed based on the neighbor structure relationships of nodes in odd and even convolutional layers.

[0054] an optimization module configured to obtain a training set by cross-pair sampling based on the dataset; calculate a loss function for the main task based on the training set and the original view; calculate a loss function for the proxy task based on heterogeneous and homogeneous neighbor comparison views; and obtain a final node representation by jointly optimizing the loss functions of the main task and the proxy task;

[0055] The item recommendation module is used to predict the relevance score between users and items based on the final node representation, and to filter out an unbiased item recommendation list for users based on the relevance score.

[0056] The third object of the present invention can be achieved by adopting the following technical solutions:

[0057] An electronic device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the above-mentioned popularity deviation elimination recommendation method is implemented.

[0058] The fourth object of the present invention can be achieved by adopting the following technical solutions:

[0059] A storage medium stores a program, which, when executed by a processor, implements the above-mentioned popularity deviation elimination recommendation method.

[0060] The present invention has the following beneficial effects compared to the prior art:

[0061] 1. This invention effectively adjusts the node representation distribution through homogeneous and heterogeneous neighbor comparison views in proxy tasks, enabling the debiased recommendation model to learn evenly distributed user and item representation vectors, thereby achieving fairer relevance score predictions and providing recommendation lists that better meet user needs.

[0062] 2. The present invention innovatively introduces cross-pairing sampling in the training process of the self-supervised recommendation model, so that the main task of the debiased recommendation model can achieve debiasing at the level of training samples, ensuring the model's excellent performance in recommendation accuracy, debiasing ability and training efficiency; the proxy task achieves debiasing at the representation level based on the main task. The two debiasing levels are organically combined to promote the model to better eliminate popularity bias in recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0064] Figure 1 This is a simplified flowchart of a popularity bias elimination recommendation method based on odd-even layer graph comparative learning and cross-pairing sampling according to Example 1 of the present invention;

[0065] Figure 2 Detailed flow chart of the popularity bias elimination recommendation method based on odd-even layer graph comparative learning and cross-pairing sampling according to Example 1 of the present invention;

[0066] Figure 3 A schematic diagram of generating heterogeneous and homogeneous neighbor comparison views in the even-odd convolutional layer of a noise-perturbed graph convolutional network according to Example 1 of the present invention;

[0067] Figure 4 This is a framework diagram of a popularity bias elimination recommendation system based on odd-even layer graph comparative learning and cross-pairing sampling according to Example 2 of the present invention;

[0068] Figure 5 This is a structural block diagram of an electronic device according to embodiment 3 of the present invention. DETAILED DESCRIPTION

[0069] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. It should be understood that the specific embodiments described are only used to explain this application and are not used to limit this application.

[0070] This embodiment provides a popularity bias elimination recommendation method based on odd-even layer graph contrast learning and cross-pairing sampling, called CrossGCL (Cross-pairwise Graph Contrastive Learning). This method uses the cross-pairing sampling method to obtain training samples, and on the basis of the unbiased ranking learning of the main task, uses the proxy task to assist in correcting the representation distribution of users and items, thereby obtaining higher-quality user and item feature information. Unlike the existing partial graph comparison method, this embodiment does not perform structural enhancement through random discarding when constructing the comparison view, but instead constructs a representation-level comparison view based on the different neighborhood relationships contained in the odd-even graph convolution layer. The objective functions of the main task and the comparison task are jointly trained using the gradient backpropagation method, so that the supervisory signal mined by the comparison task can effectively correct the learning results of the main task, thereby eliminating popularity bias in the recommendation results.

[0071] Example 1:

[0072] like Figure 1 、 2 As shown, the popularity bias elimination recommendation method based on odd-even layer graph comparative learning and cross-pairing sampling provided in this embodiment includes the following steps:

[0073] S101. Construct a bipartite graph based on the acquired dataset of user-item interactions.

[0074] In this embodiment, the public movie datasets Movielens-1M and Netflix and the fashion dataset Alibaba-iFashion are collected for training and testing the model of this embodiment.

[0075] The set of user and item interactions contained in the dataset is denoted as Each element I t =(u,i) represents a historical interaction between user u and item i.

[0076] Users and items are regarded as two types of nodes in a bipartite graph, and an interaction is regarded as an edge between two nodes to construct a bipartite graph (user-item interaction graph).

[0077] Bipartite graph In the node set Contains user node set V u and the commodity node set V i , edge set Represents the connection relationship between user and product nodes.

[0078] S102. Input the adjacency matrix of the bipartite graph into the graph convolutional network in the debiased recommendation model, and use the output mean of all graph convolutional layers as the original view of users and items.

[0079] like Figure 2 As shown, the debiased recommendation model in this embodiment includes a graph convolutional network and a noise-perturbed graph convolutional network.

[0080] Furthermore, step S102 includes:

[0081] (1) Input the adjacency matrix and initial eigenvector matrix of the bipartite graph into the graph convolutional network.

[0082] The user's initial eigenvector in the initial eigenvector matrix and the initial feature vector of the item As the output of the 0th convolutional layer.

[0083] The graph convolution in this embodiment includes three convolution layers, and the graph convolution operation is performed using three convolution layers. The graph convolution operation formula is as follows:

[0084]

[0085] Among them, the superscript l indicates the graph convolution layer where the vector is located. represents the set of neighbor nodes of user node u, Represents the set of neighbor nodes of item node i; symmetric normalization term Follow the design of standard GCN; They represent the feature vectors of users and items output by the l-th convolutional layer, l = 0, 1, 2 respectively.

[0086] (2) The output mean of all graph convolutional layers is taken as the original view of users and items. The formula is as follows:

[0087]

[0088] Among them, L is the number of graph convolution layers, e u and e i are the original views of users and items respectively.

[0089] S103. Input the adjacency matrix of the bipartite graph into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer; construct heterogeneous and homogeneous neighbor comparison views based on the neighbor structure relationship of nodes in the odd and even layers of the graph convolution.

[0090] Further, such as Figure 3 As shown, step S103 includes:

[0091] (1) The adjacency matrix and initial eigenvector matrix of the bipartite graph are input into the noise-perturbed graph convolutional network.

[0092] The user's initial eigenvector in the initial eigenvector matrix and the initial feature vector of the item As the output of the 0th convolutional layer.

[0093] The graph convolution operation is first performed using three noise perturbation convolution layers. The formula is as follows:

[0094]

[0095]

[0096] After noise disturbance, the formula is as follows:

[0097]

[0098] Where, l = 0, 1, 2, m = 0, 1, 2, 3; and are the enhanced user and item views obtained by perturbation of the m-th convolutional layer, and is the noise applied to the user and item nodes in the mth convolutional layer, f norm (·) normalizes the random noise vector ξ to a unit vector, and sign(·) restricts the direction of the normalized noise vector to be (or ) In the same hyperspace quadrant, the constant parameter δ is used to limit the length of the noise;

[0099] Normalize the augmented user and item views separately:

[0100]

[0101] in, and is the normalized representation of the node.

[0102] (2) The neighboring nodes of the node in the odd-numbered layers are regarded as heterogeneous neighbors, and the outputs of the first and third noise perturbation convolutional layers are selected to construct a heterogeneous neighbor comparison view:

[0103]

[0104] in, and are heterogeneous neighbor comparison views of users and items respectively; V u and V i are the user node set and the item node set respectively; τ is the temperature coefficient hyperparameter;

[0105] (3) Consider the neighboring nodes of the node in the even-numbered layers as isomorphic neighbors, and select the outputs of the zeroth and second noise perturbation convolutional layers to construct a comparative view of isomorphic neighbors:

[0106]

[0107]

[0108] in, and are the isomorphic neighbor comparison views of users and items respectively; V u and V i They are respectively the user node set and the item node set; and is the normalized representation of the node, and τ is the temperature coefficient hyperparameter.

[0109] S104. Based on the dataset, a training set is obtained by cross-pairing sampling; based on the training set and the original view, the loss function of the main task is calculated; based on the heterogeneous and homogeneous neighbor comparison views, the loss function of the proxy task is calculated; and the final node representation is obtained by jointly optimizing the two loss functions.

[0110] Furthermore, step S104 includes:

[0111] (1) Cross-pair sampling of all user-item interactions to obtain a training sample set Where k is the number of interactions contained in each sample, batch is the training batch, and multiple k can be used to combine different sample sets.

[0112] The loss function of the main task is calculated based on the training sample set and the original view. The formula is as follows:

[0113]

[0114] in, is the correlation score between user u and item i calculated by vector inner product, σ(·) is the sigmoid function;

[0115] (2) Based on InfoNCE and the heterogeneous neighbor comparison view, the heterogeneous neighbor comparison loss function is calculated as follows:

[0116]

[0117] (3) The homogeneous neighbor comparison loss function is calculated based on InfoNCE and homogeneous neighbor comparison view. The formula is as follows:

[0118]

[0119] (4) The total loss function of the proxy task is composed of the heterogeneous neighbor contrast loss function and the homogeneous neighbor contrast loss function, and the formula is as follows:

[0120]

[0121] Among them, the hyperparameter α is used to balance the weights of the heterogeneous neighbor contrast loss function and the homogeneous neighbor contrast loss function;

[0122] (5) Through joint training, the main task loss function and the proxy task loss function are optimized. The final optimization goal is as follows:

[0123]

[0124] Where λ1 and λ2 are regularization parameters, Θ represents the model parameters, and ‖·‖2 is the L2 regularization function. Optimizing L yields the optimal node representation.

[0125] Throughout the training process, the trainable parameters (initial feature vectors of users and items) are updated and optimized using loss backpropagation and gradient descent methods.

[0126] In this embodiment, the early stopping method is used, and the training is terminated when the NDCG (normalized discounted cumulative gain) indicator on the test set does not improve for 10 consecutive epochs.

[0127] After the model is trained, the initial feature vectors of users and items in the corresponding optimized initial feature vector matrix are is the final node representation.

[0128] S105: Predict the relevance scores between the user and the item based on the final node representation, and filter out an unbiased item recommendation list for the user based on the relevance scores.

[0129] Based on the final node representation, the relevance score of each user-item pair is calculated; for each user, the relevance scores are sorted from high to low, and the top 20 items are used as the user's item recommendation list.

[0130] The bipartite graph and the final node representation in step S101 are input into the graph convolutional network in the debiased recommendation model to obtain the original views of the user and the item as the output of the debiased recommendation model; based on the original views of the user and the item, the predicted relevance score of each item for a specific user is calculated.

[0131] Those skilled in the art will appreciate that all or part of the steps in the method for implementing the above embodiments may be completed by instructing related hardware through a program, and the corresponding program may be stored in a computer-readable storage medium.

[0132] It should be noted that although the method operations of the above embodiments are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order, or that all of the illustrated operations must be performed to achieve the desired results. Rather, the depicted steps may be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into a single step, and / or a single step may be broken down into multiple steps.

[0133] Example 2:

[0134] like Figure 4 As shown, this embodiment provides a popularity bias elimination recommendation system based on odd-even layer graph contrast learning and cross-pair sampling, which includes a construction module 401, an original view generation module 402, a contrast view generation module 403, an optimization module 404, and an item recommendation module 405, wherein:

[0135] Construction module 401, for constructing a bipartite graph based on the acquired dataset of user-item interactions;

[0136] The original view generation module 402 is used to input the adjacency matrix of the bipartite graph into the graph convolutional network in the debiased recommendation model, and use the output mean of all convolutional layers as the original view of users and items;

[0137] The comparison view generation module 403 is configured to input the adjacency matrix of the bipartite graph into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer; based on the enhanced user and item views, heterogeneous and homogeneous neighbor comparison views are constructed based on the neighbor structure relationships of nodes in odd and even convolutional layers.

[0138] Optimization module 404 is configured to obtain a training set by cross-pair sampling based on the dataset; calculate a loss function for the primary task based on the training set and the original view; calculate a loss function for the proxy task based on heterogeneous and homogeneous neighbor comparison views; and obtain a final node representation by jointly optimizing the loss functions of the primary and proxy tasks.

[0139] The item recommendation module 405 is used to predict the relevance score between the user and the item based on the final node representation, and filter out an unbiased item recommendation list for the user based on the relevance score.

[0140] The specific implementation of each module in this embodiment can be found in the above-mentioned embodiment 1, and will not be described one by one here; it should be noted that the system provided in this embodiment is only illustrated by the division of the above-mentioned functional modules. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0141] Example 3:

[0142] This embodiment provides an electronic device, which may be a computer, such as Figure 5 As shown, a processor 502, a memory, an input device 503, a display 504, and a network interface 505 are connected via a system bus 501. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium 506 and an internal memory 507. The non-volatile storage medium 506 stores an operating system, a computer program, and a database. The internal memory 507 provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the processor 502 executes the computer program stored in the memory, the method for eliminating popularity deviation of the recommendation system of the above-mentioned embodiment 1 is implemented as follows:

[0143] Construct a bipartite graph based on the obtained dataset of user-item interactions;

[0144] The adjacency matrix and initial eigenvector matrix of the bipartite graph are input into the graph convolutional network in the debiased recommendation model, and the output mean of all convolutional layers is used as the original view of users and items;

[0145] The adjacency matrix of the bipartite graph is input into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer. Based on the enhanced user and item views, heterogeneous and homogeneous neighbor comparison views are constructed based on the neighbor structure relationships of nodes in odd and even convolutional layers.

[0146] Based on the dataset, a training set is obtained by cross-pair sampling; the loss function of the main task is calculated based on the training set and the original view; the loss function of the proxy task is calculated based on the heterogeneous and homogeneous neighbor comparison views; and the final node representation is obtained by jointly optimizing the loss functions of the main task and the proxy task.

[0147] The relevance scores of users and items are predicted based on the final node representation, and an unbiased item recommendation list is filtered out for the user based on the relevance scores.

[0148] Example 4:

[0149] This embodiment provides a storage medium, which is a computer-readable storage medium and stores a computer program. When the computer program is executed by a processor, the method for eliminating popularity deviation in the recommendation system of the above-mentioned embodiment 1 is implemented as follows:

[0150] Construct a bipartite graph based on the obtained dataset of user-item interactions;

[0151] The adjacency matrix and initial eigenvector matrix of the bipartite graph are input into the graph convolutional network in the debiased recommendation model, and the output mean of all convolutional layers is used as the original view of users and items;

[0152] The adjacency matrix of the bipartite graph is input into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer. Based on the enhanced user and item views, heterogeneous and homogeneous neighbor comparison views are constructed based on the neighbor structure relationships of nodes in odd and even convolutional layers.

[0153] Based on the dataset, a training set is obtained by cross-pair sampling; the loss function of the main task is calculated based on the training set and the original view; the loss function of the proxy task is calculated based on the heterogeneous and homogeneous neighbor comparison views; and the final node representation is obtained by jointly optimizing the loss functions of the main task and the proxy task.

[0154] The relevance scores of users and items are predicted based on the final node representation, and an unbiased item recommendation list is filtered out for the user based on the relevance scores.

[0155] It should be noted that the computer-readable storage medium of the present embodiment may be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0156] In summary, the popularity bias elimination recommendation method, system, electronic device and storage medium provided by the present invention based on odd-even layer graph contrast learning and cross-pairing sampling construct a "user-item interaction bipartite graph" according to the interaction data between users and items in the data set; the bipartite graph adjacency matrix and the initial node feature vector are respectively input into the graph convolution network and the noise-perturbed graph convolution network in the debiased recommendation model to perform neighbor information aggregation, and aggregate the information of the neighbor nodes of the user node or item node to its own node to realize the learning of local graph structural features; wherein, the graph convolution network outputs the mean of the convolution layer outputs of different convolution times as the final node vector; the noise-perturbed graph convolution network regards the neighboring nodes of the node in the odd layer as heterogeneous neighbors, The output of the noise perturbation convolution layer where the odd number is located is selected to construct a heterogeneous neighbor comparison view; the neighboring nodes of the node in the even number layer are regarded as homogeneous neighbors, and the output of the noise perturbation convolution layer where the even number is located is selected to construct a homogeneous neighbor comparison view; based on the different neighborhood relationships contained in the odd-even graph convolution layer, a representation-level comparison view is constructed, and effective supervision signals are mined in a fine-grained manner in the proxy task to assist in guiding the main task to perform ranking learning; the main task loss function is calculated using the training batch data obtained by cross-pairing sampling; the main task and proxy task loss functions are jointly optimized to obtain the optimal user and item representation vectors; the user and item relevance scores are predicted based on the optimal user and item representation vectors, and the item recommendation list for the specified user is obtained based on the predicted relevance scores. By constructing a graph convolution layer comparison view based on different neighborhood relationships in the proxy task, the present invention can mine massive amounts of unlabeled data, correct the node representations learned by the unbiased ranking task of the main task, obtain a fairer predicted relevance score, and effectively alleviate the popularity bias in the recommendation results.

[0157] The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes based on the technical solution and inventive concept of the present invention within the scope disclosed by the present invention, which falls within the scope of protection of the present invention.

Claims

1. A popularity bias elimination recommendation method based on odd-even layer graph contrast learning and cross-pair sampling, characterized in that: The method comprises: Construct a bipartite graph based on the obtained dataset of user-item interactions; The adjacency matrix and initial eigenvector matrix of the bipartite graph are input into the graph convolutional network in the debiased recommendation model, and the output mean of all convolutional layers is used as the original view of users and items; The adjacency matrix of the bipartite graph is input into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer; based on the enhanced user and item views, heterogeneous and homogeneous neighbor comparison views are constructed based on the neighbor structure relationships of nodes in odd and even convolutional layers, respectively; the noise-perturbed graph convolutional network has three noise-perturbed convolutional layers; Based on the dataset, a training set is obtained by cross-pair sampling; the loss function of the main task is calculated based on the training set and the original view; the loss function of the proxy task is calculated based on the heterogeneous and homogeneous neighbor comparison views; and the final node representation is obtained by jointly optimizing the loss functions of the main task and the proxy task. Predict the relevance score between users and items based on the final node representation, and filter out an unbiased item recommendation list for the user based on the relevance score; The enhanced user and item views are based on the neighbor structure relationships of nodes in odd and even convolution layers to construct heterogeneous and homogeneous neighbor comparison views, respectively, including: The neighboring nodes of a node in an odd-numbered layer are regarded as heterogeneous neighbors, and the output of the noise-perturbed convolutional layer where the odd number is located is selected to construct a heterogeneous neighbor comparison view; The neighboring nodes of a node in the even-numbered layer are regarded as isomorphic neighbors, and the output of the noise-perturbed convolutional layer where the even number is located is selected to construct a comparative view of isomorphic neighbors.

2. The popularity deviation elimination recommendation method according to claim 1, characterized in that: The neighboring nodes of the node in the odd-numbered layer are regarded as heterogeneous neighbors, and the output of the noise perturbation convolution layer where the odd number is located is selected to construct a heterogeneous neighbor comparison view, including: Where, and The heterogeneous neighborhood comparison views of users and items respectively; V u and V i are the user node set and the item node set respectively; τ is the temperature coefficient hyperparameter; and are the normalized representations of user nodes and item nodes respectively, and are the enhanced user view and item view output by the mth noise perturbation convolutional layer, m = 0, 1, 2, 3; The neighboring nodes of the node in the even-numbered layer are regarded as isomorphic neighbors, and the output of the noise perturbation convolution layer where the even-numbered nodes are located is selected to construct a isomorphic neighbor comparison view, including: Where, and Isomorphic neighborhood comparison views for users and items, respectively.

3. The popularity deviation elimination recommendation method according to any one of claims 1 to 2, characterized in that: The adjacency matrix of the bipartite graph is input into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer, including: The adjacency matrix and initial eigenvector matrix of the bipartite graph are input into the noise-perturbed graph convolutional network. The graph convolution operation is first performed using three noise-perturbed convolutional layers. The formula is as follows: Where l = 0, 1, 2; are the feature vectors of users and items output by the l+1th convolutional layer respectively; is the set of neighbor nodes of user node u, is the set of neighbor nodes of item node i; symmetric normalization term Follow the design of standard GCN; are the feature vectors of users and items output by the lth convolutional layer respectively; and The initial feature vectors of users and items in the initial feature vector matrix are used as the output of the 0th convolutional layer; After noise disturbance, the formula is as follows: Where m = 0, 1, 2, 3; and are the enhanced user and item views obtained by perturbation of the m-th convolutional layer; and is the noise applied to the user and item nodes in the mth convolutional layer; f norm (·) normalizes the random noise vector ξ to a unit vector; sign(·) restricts the direction of the normalized noise vector to (or ) in the same hyperspace quadrant; the constant parameter δ is used to limit the length of the noise.

4. The popularity deviation elimination recommendation method according to claim 1, characterized in that: The final node representation is obtained by jointly optimizing the loss functions of the main task and the proxy task, including: According to the loss function of the main task and the loss function of the proxy task, the optimization objective is as follows: Where, The loss function of the main task; is the loss function of the proxy task; λ1 and λ2 are regularization parameters; Θ represents the model parameters, and ‖·‖2 is the l2 regularization function; During the training process, the loss back propagation and gradient descent methods are used to optimize That is, the final node representation is obtained.

5. The popularity deviation elimination recommendation method according to any one of claims 1 and 4, characterized in that: Cross-pair sampling is performed on all interactions between users and items in the dataset to obtain a training set Where k is the number of interactions contained in each sample; b is the training batch; u s and i s are user nodes and product nodes respectively, s=1,2,...,bk; different samples are combined by taking multiple k; The formula of the loss function of the main task is as follows: Where σ(·) is the sigmoid function; is the correlation score between the user and the item calculated by the vector inner product; e u and e i are the original views of users and items respectively.

6. The popularity deviation elimination recommendation method according to any one of claims 1 and 4, characterized in that: The formula of the loss function of the proxy task is as follows: Where, is the homogeneous neighbor contrast loss function calculated based on InfoNCE and heterogeneous neighbor contrast views, is the heterogeneous neighbor contrast loss function calculated based on InfoNCE and homogeneous neighbor contrast views; the hyperparameter α is used to balance the weights of the heterogeneous neighbor contrast loss function and the homogeneous neighbor contrast loss function.

7. The popularity deviation elimination recommendation method according to any one of claims 1 to 2 and 4, characterized in that: The step of constructing a bipartite graph based on the acquired dataset of user-item interactions includes: The set of user and item interactions contained in the dataset is denoted as Each element I t =(u,i) is a historical interaction between user u and item i; Consider users and items as two types of nodes in a bipartite graph, and an interaction as an edge between two nodes to construct a bipartite graph. Node Collection Contains user node set V u and the commodity node set V i ; Edge set It is the connection relationship between user nodes and product nodes.

8. A popularity bias elimination recommendation system based on odd-even layer graph contrast learning and cross-pair sampling, characterized by: The system comprises: A construction module is used to construct a bipartite graph based on the obtained dataset of user-item interactions; The original view generation module is used to input the adjacency matrix of the bipartite graph into the graph convolutional network in the debiased recommendation model and use the mean output of all convolutional layers as the original view of users and items; A comparative view generation module is configured to input the adjacency matrix of the bipartite graph into the noise-perturbed graph convolutional network in the debiased recommendation model to obtain enhanced user and item views output by each noise-perturbed convolutional layer; based on the enhanced user and item views, heterogeneous and homogeneous neighbor comparative views are constructed based on the neighbor structure relationships of nodes in odd and even convolutional layers, respectively; the noise-perturbed graph convolutional network has three noise-perturbed convolutional layers; an optimization module configured to obtain a training set by cross-pair sampling based on the dataset; calculate a loss function for the main task based on the training set and the original view; calculate a loss function for the proxy task based on heterogeneous and homogeneous neighbor comparison views; and obtain a final node representation by jointly optimizing the loss functions of the main task and the proxy task; The item recommendation module is used to predict the relevance score between users and items based on the final node representation, and filter out an unbiased item recommendation list for users based on the relevance score; The enhanced user and item views are based on the neighbor structure relationships of nodes in odd and even convolution layers to construct heterogeneous and homogeneous neighbor comparison views, respectively, including: The neighboring nodes of a node in an odd-numbered layer are regarded as heterogeneous neighbors, and the output of the noise-perturbed convolutional layer where the odd number is located is selected to construct a heterogeneous neighbor comparison view; The neighboring nodes of a node in the even-numbered layer are regarded as isomorphic neighbors, and the output of the noise-perturbed convolutional layer where the even number is located is selected to construct a comparative view of isomorphic neighbors.

9. A storage medium storing a program, characterized in that: When the program is executed by a processor, the popularity deviation elimination recommendation method according to any one of claims 1 to 7 is implemented.

Images