Popularity bias elimination intelligent recommendation method based on double-branch graph neural network

Abstract

The application discloses a kind of based on double branch graph neural network's popularity bias removal intelligent recommendation method, the method obtains original data set;Delete the user item exchange record of popular item in the original data set, obtain long tail data set;Based on the original data set and the long tail data set, original subgraph and long tail subgraph are respectively constructed;For the original subgraph and the long tail subgraph, information propagation is carried out by graph neural network respectively, and in the training process, the weight of the corresponding graph neural network of long tail subgraph is gradually increased based on a dynamically adjusted parameter, to obtain the item recommendation result of user based on the corresponding graph neural network of long tail subgraph after training, the popularity bias problem caused by data obeying long tail distribution in traditional recommendation scene is solved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and data mining, and in particular relates to a popularization-biased intelligent recommendation method based on a dual-branch graph neural network. Background Technology

[0002] With the booming development of online information, recommender systems are playing an increasingly important role in the internet ecosystem. Recommender systems are a core user-facing technology in the internet age, aiming to solve the problem of data explosion by recommending small-scale personalized data to users. Traditional collaborative filtering-based recommender algorithms analyze users' interaction data to determine their potential preferences.

[0003] Popularity bias refers to the phenomenon where popular items are recommended significantly more frequently than their actual popularity suggests. Also known as the long-tail problem, this phenomenon is common in recommendation data: in most cases, a small subset of popular items accounts for the majority of user interactions. When recommendation systems model on such long-tail data, the model often assigns higher scores to popular items than expected and simply treats less popular items as negative examples. Therefore, popular items are recommended even more frequently than their initial popularity was initially indicated in the dataset.

[0004] Currently, the more mature recommendation methods to remove popularity bias mainly include regularization, adversarial learning, and rebalancing. Regularization-based methods push the model towards a more balanced recommendation list through appropriate regularization. Early methods used Laplace regularization to balance item popularity. In information-neutral recommendation systems, mean-matching regularization is used to correct popularity bias. Procedural regularization methods attribute the ineffective recommendation of long-tail items to insufficient training of the model on long-tail items. A method called Full Space Adaptation Model (ESAM) approaches this from the perspective of domain adaptation, introducing three regularization terms to transfer knowledge from well-trained popular items to long-tail items. Although regularization-based methods have achieved some success, the constraint of manually designed regularization terms makes it difficult for the model to adapt to large-scale and real-time recommendation scenarios. The basic idea of ​​reinforcement learning-based methods is to conduct a minimum-maximum game between the recommender G and the introduced adversarial D, causing D to give a signal to increase the recommendation opportunity for long-tail items. Rebalancing-based methods first generate a relatively large recommendation list using traditional recommendation models, and then reorder the list by considering item popularity. A common reordering strategy is to introduce propensity scores, and the three methods mentioned above can all improve the popularity bias problem to some extent. However, regularization-based methods require manually setting regularization terms, and the effectiveness of these terms is often closely related to the data distribution, lacking generalization ability and making them difficult to apply to large-scale dynamic data scenarios. Reinforcement learning-based methods introduce noise, which significantly affects the overall recommendation accuracy. Rebalancing methods typically require human experience or substantial computational resources to select confidence weights or propensity scores. Summary of the Invention

[0005] Based on the above-mentioned development status, the purpose of this invention is to design a popularity bias-free intelligent recommendation method based on a dual-branch graph neural network to solve the popularity bias problem caused by the long-tail distribution of data in traditional recommendation scenarios.

[0006] The technical content of this invention includes:

[0007] A method for intelligent recommendation without popularity bias based on a bi-branch graph neural network, characterized in that the method includes:

[0008] Obtain the original dataset; the original dataset contains several user item exchange records;

[0009] Remove user item exchange records for popular items from the original dataset to obtain a long-tail dataset;

[0010] Based on the original dataset and the long-tail dataset, an original subgraph and a long-tail subgraph are constructed respectively, wherein the nodes in the original subgraph and the long-tail subgraph include: user nodes and item nodes;

[0011] For the original subgraph and the long-tail subgraph, information is propagated through graph neural networks. During the training process, the weight of the graph neural network corresponding to the long-tail subgraph is gradually increased based on a dynamically adjusted parameter, so as to obtain the user's item recommendation results based on the graph neural network corresponding to the trained long-tail subgraph.

[0012] Furthermore, by removing user item exchange records for popular items from the original dataset, a long-tail dataset is obtained, including:

[0013] Based on the Pareto principle, user item exchange records for popular items are removed from the original dataset to obtain a long-tail dataset.

[0014] Further, constructing the original subgraph based on the original dataset includes:

[0015] Set the first-order neighbors of the user node as item nodes with exchange records;

[0016] Set the second-order neighbors of the user node to user nodes that have the same exchanged items;

[0017] Set the third-order neighbors of the user node to the item nodes that have exchange records with the second-order neighbors;

[0018] The user node's third-order and higher neighbors are set as high-order cooperative signals.

[0019] Furthermore, constructing a long-tail subgraph based on the long-tail dataset includes:

[0020] Based on the user item exchange records in the long-tail dataset, an initial long-tail subgraph is constructed; wherein, the nodes in the initial long-tail subgraph include: user nodes and item nodes, the first-order neighbors of the user node are item nodes with exchange records, the second-order neighbors of the user node are user nodes with the same exchanged item, the third-order neighbors of the user node are item nodes with exchange records with the second-order neighbors, and the third-order and above neighbors of the user node are higher-order cooperative signals.

[0021] The similarity between user nodes is calculated using the point-to-point mutual information method;

[0022] Based on the similarity, edges are added between user nodes in the initial long-tailed subgraph to obtain the long-tailed subgraph.

[0023] Further, adding edges between user nodes in the initial long-tail subgraph based on the similarity includes:

[0024] Calculate the interaction probability between user u1 and items in the item set S of the long-tail dataset. Where |S| represents the number of items in the item set S, and |S(u1)| represents the number of interaction items between user u1 and the items in the item set S;

[0025] Calculate the interaction probability between user u2 and items in the item set S of the long-tail dataset. Where |S(u2)| represents the number of interaction items between user u2 and items in the item set S;

[0026] Calculate the probability that user u1 and user u2 have the same interaction item in the item set S.

[0027] Based on the interaction probability p(u1), the interaction probability p(u2), and the probability p(u1, u2), the PMI value PMI(u1, u2) is obtained;

[0028] If the PMI value is greater than zero, a weighted edge is added between the nodes corresponding to user u1 and user u2, and the PMI value is used as the weight of the weighted edge.

[0029] Furthermore, for the original subgraph and the long-tail subgraph, information is propagated through graph neural networks respectively. During training, the weights of the graph neural network corresponding to the long-tail subgraph are gradually increased based on a dynamically adjusted parameter, so as to obtain the user's item recommendation results based on the trained graph neural network corresponding to the long-tail subgraph, including:

[0030] Obtain the true labels of the original data;

[0031] The graph neural network corresponding to the original subgraph performs information propagation in the Tth round to obtain the first predicted label;

[0032] The graph neural network corresponding to the long-tailed subgraph performs information propagation in the Tth round to obtain the second prediction.

[0033] Based on the first predicted label and the true label, calculate the T-th round loss function of the graph neural network corresponding to the original subgraph.

[0034] Based on the second predicted label and the true label, calculate the T-th round loss function of the graph neural network corresponding to the original subgraph.

[0035] Calculate the parameters for round T. Among them, T max Indicates the total number of training rounds;

[0036] Based on the total loss function of round T Adjust the parameters of the graph neural network corresponding to the original subgraph and the graph neural network corresponding to the long-tailed subgraph;

[0037] When the number of training rounds reaches the total number of training rounds T max At that time, the graph neural network corresponding to the trained long-tailed subgraph is obtained;

[0038] Based on the graph neural network corresponding to the trained long-tailed subgraph, the item recommendation results for the user are obtained.

[0039] Further, the step of performing the Tth round of information propagation in the graph neural network corresponding to the original subgraph to obtain the first predicted label includes:

[0040] The vector representation e of any node e in the original subgraph l-1 By aggregating the vector representations of neighboring nodes, we obtain the vector representation e. l Where l represents the number of aggregations, l = 0, 1, ..., L-1, L represents the number of layers in the graph neural network corresponding to the original subgraph, and node e represents a user node or an item node;

[0041] According to the vector representation e l This yields the implicit representation H of each node at each level. l ;

[0042] Based on user historical interaction data, a user-item interaction matrix R is generated to obtain the adjacency matrix A of the graph neural network corresponding to the original subgraph;

[0043] Based on the implicit representation H l Perform multi-level forward propagation with the adjacency matrix A to obtain the hidden representation H. l+1 =δ(D -1 / 2 AD -1 / 2 H l W l ); where D represents the angle matrix, δ represents the nonlinear activation function of the graph neural network, and W represents the weights;

[0044] According to the implicit representation H L-1 Obtain the first predicted label.

[0045] A de-popularity biased intelligent recommendation device based on a dual-branch graph neural network, the device comprising:

[0046] The data acquisition module is used to acquire the raw dataset; the raw dataset contains several user item exchange records;

[0047] The preprocessing module is used to remove user item exchange records of popular items from the original dataset to obtain a long-tail dataset;

[0048] The graph construction module is used to construct an original subgraph and a long-tail subgraph based on the original dataset and the long-tail dataset, respectively, wherein the nodes in the original subgraph and the long-tail subgraph include user nodes and item nodes;

[0049] The intelligent recommendation module is used to propagate information through graph neural networks for both the original subgraph and the long-tail subgraph. During training, the weights of the graph neural network corresponding to the long-tail subgraph are gradually increased based on a dynamically adjusted parameter. Based on the trained graph neural network corresponding to the long-tail subgraph, the module obtains item recommendation results for the user.

[0050] A computer device, characterized in that it includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform any of the methods described above.

[0051] A storage medium storing a computer program, wherein the computer program is configured to execute any of the methods described above when run.

[0052] Compared with existing technologies, this invention solves the popularity bias problem in recommendation systems and improves the recommendation rate of long-tail items. This method uses the Pareto principle to remove popular items and constructs a long-tail subgraph to model the features of long-tail items separately. The long-tail subgraph with lower overall node degree gives the graph neural network a larger receptive field, allowing the model to move beyond focusing on locally popular items and aggregate more long-tail item information with fewer hops. This invention proposes an innovative dual-branch graph neural network. On the one hand, it models the user's basic characteristics based on the original data distribution. On the other hand, it focuses on mining long-tail item features from the subgraph. The ratio of the two branches is dynamically adjusted through an accumulative learning module, gradually shifting the overall attention towards long-tail items. Attached Figure Description

[0053] Figure 1 This is a subgraph construction diagram of the long-tail branch in this invention;

[0054] Figure 2 This is a model architecture diagram of the present invention;

[0055] Figure 3 This is a flowchart of the present invention. Detailed Implementation

[0056] To make the objectives, solutions, and advantages of this invention clearer, an experiment conducted on a real dataset is used as an example to further illustrate the invention in detail. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0057] This invention presents a BiGNN-based intelligent recommendation method for removing popularity bias. Addressing the problem of repeated recommendations of a small number of popular items that constitute the majority of interaction records, and the inability to effectively train and mine long-tail items, this invention considers appropriately reducing the weight of popular items while maintaining recommendation accuracy, and gradually shifting the model's learning objective towards long-tail items to achieve the goal of mining more long-tail items. Furthermore, this invention discloses a bi-branch network based on cumulative learning, where the original data branch and the long-tail branch (excluding popular items) share parameters for cumulative learning, allowing the model's learning objective to naturally transition towards long-tail items. Popular items in the original data are reduced in weight through degree penalty to weaken their impact on the recommendation results. Additionally, for the long-tail branch (excluding popular items), to address the problem of impaired graph connectivity due to missing data, a user similarity calculation based on point relevance scores is designed, and the similarity score is used as the weight to add weighted edges between users.

[0058] Furthermore, this invention also considers that in the user-item interaction bipartite graph, the excessive number of edges connecting popular item nodes can cause the graph neural network to be trapped around the locally dense popular item nodes during graph knowledge aggregation, making it difficult to aggregate information from long-tail nodes. Taking movie recommendation as an example, some popular movies have high click-through rates, causing the model to tend to recommend popular items to users, while some less-watched gems remain undiscovered. To fully explore long-tail items, this method uses the Pareto principle to remove popular items, resulting in a long-tail subgraph containing only low-degree nodes. Considering the connectivity of the long-tail subgraph, and based on the assumption that users with similar features may have the same preferences, this method introduces the Pointwise Mutual Information (PMI) method to increase the number of edges between users by calculating similarity. Then, a graph neural network (GNN) model is used to aggregate neighborhood information layer by layer from the subgraph. Another branch of this method's model is called the regular learning branch. This method uses the same GNN structure as the previous long-tail branch, but the input data for this branch is a user-item interaction graph that follows the original data distribution. The function of this branch is to obtain the global features of users by learning the original data distribution. The two branches are connected via a cumulative learning module. The entire BiGNN model is adjusted to first learn general features from the original distribution, and then gradually focus on the tail data. In particular, the graph neural networks of the two branches share parameters, making the traditional learning branches more effective for learning the long-tail branches.

[0059] The implementation steps of this plan are as follows:

[0060] 1) Analyze the original data distribution to distinguish between popular items with many interaction records and long-tail items with few interaction records;

[0061] 2) Based on the characteristics of the bipartite graph neural network model, the input data is modeled as a graph structure. For the regular learning branch, the input data is a bipartite graph of user-item interactions based on the original dataset;

[0062] 3) On another branch, the goal of this invention is to eliminate the popularity bias of popular items and focus on modeling the characteristics of long-tail items. To achieve this, this invention follows the Pareto principle, that is, 20% of popular items account for the majority of the overall interaction records. Through statistical analysis of the dataset, the top 20% of popular items in terms of total interaction frequency are removed, and the remaining data is used as the input data for the long-tail branch.

[0063] 4) For both branches, this invention establishes graph neural networks to propagate information along the user-item interaction graph and update the embeddings of users and items. The advantage of graph neural networks lies in enhancing the vector representations of users and items by aggregating neighborhood information layer by layer in the interaction graph;

[0064] 5) This invention does not process the self-connections of a node itself, but only considers the aggregation of information from other neighboring nodes that do not include the target node itself. A single GNN layer can aggregate information from the first-order nearest neighbors of the target node. Therefore, this method can naturally aggregate multi-hop neighbor information by stacking multi-layer graph neural networks, thereby more broadly and deeply exploring the user's potential interests;

[0065] 6) For long-tail subgraphs, removing popular items (i.e., deleting nodes with high degree) affects the connectivity of the subgraph to some extent. This method finds that user similarity can also be used as a feature to enhance user representation. Based on the above analysis, this method intends to use the Pointwise Mutual Information (PMI) method, widely used in Natural Language Processing (NLP), for text classification and word association. This method uses the PMI algorithm to calculate user similarity and adds weighted edges between users;

[0066] 7) In the long-tail learning branch, this method uses the same graph neural network (GNN) model as the regular representation branch;

[0067] 8) This invention introduces a cumulative learning module to combine the two branch networks. The function of the cumulative learning module is to gradually shift the model's focus towards the long-tail learning branch as the training epochs increase. To this end, this method designs a dynamically adjustable parameter to control the weights of the two networks and adjust the loss function;

[0068] 9) Throughout the training process, weights are shared between the two branch networks. This reduces the computational complexity of the model, while the embeddings learned in the traditional learning branches can help the long-tail branches. This invention uses mini-batch ADAM to optimize the prediction model and update the model parameters.

[0069] To address the popularity bias problem in recommendation systems and improve the recommendation rate of long-tail items, this invention employs the Pareto principle to remove popular items and constructs a long-tail subgraph to model the features of long-tail items separately. The long-tail subgraph, with its lower overall node degree, provides the graph neural network with a larger receptive field, allowing the model to move beyond focusing on locally popular items and aggregate more long-tail item information with fewer hops. This invention proposes an innovative dual-branch graph neural network. On one hand, it models the user's basic characteristics based on the original data distribution. On the other hand, it focuses on mining long-tail item features from the subgraph. Through an accumulative learning module, the ratio of the two branches is dynamically adjusted, gradually shifting the overall attention towards long-tail items.

[0070] Furthermore, the present invention consists of three parts:

[0071] a) Traditional learning branches acquire general features of users and items through graph neural networks;

[0072] b) The input to the long-tail branch is the user and item subgraph excluding popular items, and the PMI algorithm is used to add weighted edges between users in the long-tail subgraph. In low-degree long-tail subgraphs, the GNN has a larger receptive field, which can aggregate the information of long-tail items within a few hops, increasing the exposure rate of long-tail items;

[0073] c) The cumulative learning module combines the two branches, causing the model's attention to gradually shift towards the long tail side.

[0074] Furthermore, the aforementioned conventional branch learns users' basic preferences by modeling the original dataset. This is achieved by constructing an interaction graph of user items from the user's historical interaction data. First-order neighbors are items directly purchased by the user in their historical consumption records; second-order neighbors are users with similar purchasing behavior; and third-order neighbors are items directly purchased by second-order neighbor users. Information about neighbors of order three and above represents higher-order collaborative signals.

[0075] Furthermore, the graph neural network of the conventional branch aggregates neighborhood information layer by layer in the graph and updates the vector representations of users and items. Through these strategies, the traditional learning branch can learn the distribution characteristics of the original data well, thereby improving the ability to represent the basic features of users and items;

[0076] Furthermore, the input to the long-tail branch is first the number of interactions for all items in the data set. Inspired by the Pareto principle, which states that 20% of the most popular items account for 80% of the recorded interactions, this method removes the top 20% of items with the most interactions and uses the remaining long-tail items to construct a long-tail item-user interaction subgraph.

[0077] Furthermore, when removing popular items from the long-tail subgraph, the local density of the graph decreases and the receptive field of the GNN increases. Therefore, the model can aggregate more long-tail node information in fewer hops, significantly improving the exposure rate of long-tail items and thus eliminating popularity bias in the data.

[0078] Furthermore, since the connectivity of the long-tail subgraph is affected by the removal of popular nodes, the PMI algorithm is used to calculate the similarity between users, and weighted edges are added between users. The PMI calculation formula is as follows:

[0079]

[0080] The methods for calculating p(u1,u2) and p(u) are as follows:

[0081]

[0082]

[0083] Where S(u) represents the set of items that a user has historically interacted with, and |S(u1)∩S(u2)| is the set of items that both users u1 and u2 have interacted with. p(u1,u2) refers to the joint probability of u1 and u2, which is the probability that two users have the same interactive items; p(u) is the probability that a user interacts with items in the dataset. This invention requires setting a threshold T to remove users who are truly similar to the user. In this case, this invention only adds edges to users with a PMI score greater than 0, and uses the score as a weight to reflect the importance of different neighbors in the graph.

[0084] Furthermore, for both branches, this method establishes graph neural networks to propagate information along the user-item interaction graph and update the embeddings of users and items. The advantage of graph neural networks lies in enhancing the vector representations of users and items by aggregating neighborhood information layer by layer in the interaction graph. Taking user u as an example, the information aggregation process is as follows:

[0085]

[0086] Among them, e u The vector representation of a user node, e j The vector representation of the neighboring nodes of a user node. This represents the set of neighboring nodes of a user node, where l represents the number of information aggregations.

[0087] As a fundamental operation of graph neural networks, ACG is an aggregation function that merges the representations of the target node and its neighboring nodes. To better fit the recommendation task, this invention employs a basic weighted sum aggregator. The graph convolution operation in this invention's model is defined as follows:

[0088]

[0089]

[0090] The symmetric normalization term in this formula The aim is to avoid excessively large embedding ranges caused by graph convolution operations. Existing graph convolution operations, besides aggregating neighborhood information, also require special handling of self-connectivity issues. In this work, the invention only considers connected neighbors, not the target node itself. Instead of specifically handling self-connectivity, the invention incorporates the original node embedding as one of the same features as the output of each layer.

[0091] Furthermore, the adjacency matrix of the graph neural network is defined as follows. Given user historical interaction data, this invention can generate a user-item interaction matrix R. For each observed user-item direct interaction, this invention sets the values ​​of these user-item pairs in the matrix to 1 and their remainders to 0. Based on this, an adjacency matrix is ​​constructed:

[0092]

[0093] Furthermore, this invention utilizes multiple forward propagation layers to update the representation matrix of nodes, and the information aggregation between layers can be represented as follows:

[0094] H l+1 =δ(D -1 / 2 AD -1 / 2 H l W l ), l=0,1,…,L-1

[0095] H represents the implicit representation of each node in each layer, D is the angle matrix used to normalize the adjacency matrix A, L represents the number of stacked multi-layer graph neural networks, δ represents the non-linear activation function of the graph neural network, W represents the training weights, and L represents the number of layers in the graph neural network. A single GNN layer can aggregate information from the first-order neighbors of the target node. Therefore, this invention can naturally aggregate multi-hop neighbor information by stacking multi-layer graph neural networks, thereby more broadly and deeply exploring the user's potential interests.

[0096] Furthermore, the cumulative learning module combines the two branch networks. The function of the cumulative learning module is to gradually shift the model's focus towards the long-tail learning branch as the training epochs increase. To this end, this invention designs a dynamically adjustable parameter to control the weights of the two networks and adjust the loss function. The weight w is defined as:

[0097]

[0098] Where T max represents the total number of training epochs, while T represents the current training epoch. For each training branch, this invention uses the BPR loss as the optimization objective, that is, maximizing the distance between positive and negative samples:

[0099]

[0100] Furthermore, this invention utilizes parameter-weighted loss to obtain the final loss function:

[0101]

[0102] Furthermore, in the initial training phase of the dual-branch network, the traditional learning branch occupies a large proportion, primarily used to learn general embeddings for users and items. As w decreases, the model's attention shifts to the long-tail branch, gradually improving the recommendation rate of long-tail items. Throughout the training process, weights are shared between the two branch networks. This reduces the model's computational complexity, while the embeddings learned in the traditional learning branch aid in the learning of the long-tail branch.

[0103] Taking the movie recommendation scenario as an example, this paper describes the specific implementation steps of removing popularity bias using a dual-branch network.

[0104] Movie recommendation systems estimate user preferences based on their viewing history, ratings, and other user behaviors, and then match movies accordingly. However, in real-world movie recommendation datasets, such as Movielens, some popular movies have high click-through rates, causing the model to tend to recommend popular items to users, while some less-watched but excellent films are overlooked.

[0105] In this embodiment, the objective of this invention is to propose a bidirectional branching graph neural network to solve the long-tail item recommendation problem. The regular learning branch and the long-tail learning branch model the general characteristics of users and items, the characteristics of long-tail items, and the niche preferences of users, respectively. A cumulative learning strategy is employed to gradually shift the model's focus towards the long-tail side, thereby improving the recommendation rate of long-tail items.

[0106] This invention uses Movielen10M, a commonly used movie recommendation dataset in the recommender system field, as the experimental dataset. Movielen10M contains approximately ten million movie rating records from users on movie websites. Analysis of the data reveals a long-tail distribution, with a small number of popular items accounting for the majority of interaction records, while interaction records for less popular movies are very few. In the user-item interaction graph, popular items have more edges. The high node degree results in a locally dense graph. When a graph neural network aggregates information layer by layer in the interaction graph, it is limited by the high-degree popular nodes, making it difficult to spread to long-tail nodes.

[0107] To address the popularity bias problem in movie recommendation scenarios, i.e., to remove the negative impact of popular items on long-tail modeling, this example designs a two-branch network. One branch models the user's basic preferences on the original data distribution, while the other branch models long-tail items separately. The model structure is designed so that attention gradually shifts to the long-tail items during training.

[0108] The model in this invention consists of three parts: a) a traditional learning branch acquires general features of users and items through a graph neural network; b) the long-tail branch takes as input a user and item subgraph excluding popular items and adds weighted edges between users in the long-tail subgraph using the PMI algorithm. In low-degree long-tail subgraphs, the GNN has a larger receptive field, which can aggregate information of long-tail items within a few hops, increasing the exposure rate of long-tail items; c) the cumulative learning module combines the two branches, causing the model's attention to gradually shift towards the long-tail side. The following sections will provide a detailed introduction to the conventional representation branch, the long-tail learning branch, and the cumulative learning module.

[0109] This embodiment sets up specific experiments to verify the effectiveness of the method in removing popularity bias. Due to the special nature of long-tail item recommendations, the test sets generated by traditional time slicing or sampling methods still follow a long-tail distribution, causing the test results to not intuitively reflect the effectiveness of long-tail item recommendations. Accordingly, this method uses inverse propensity scores to rebalance the data to reduce the probability of popular items appearing, thus ensuring that items in the test set have equal probabilities. This method selects 10% of items with equal probabilities as the test set, 10% as the validation set, and the remainder as the training dataset. This method chooses two traditional evaluation metrics: recall@K and normalized loss gain ndcg@K. By default, this method sets K=20. This method sets the embedding size of all models to 64 and initializes the embedding parameters using the Xavier method.

[0110] In this embodiment, the model's output comprehensively considers both the basic preferences of users in the regular branch and the niche preferences of users in the long-tail branch. Cumulative learning enables the model to learn and extract feature knowledge from both aspects. This significantly improves the diversity of recommendation results while maintaining recommendation accuracy. The recommendation system can be viewed as a loop, where the recommendation results influence the user's next action and decision. As more long-tail movies are recommended to users, the interaction records for these items increase, and the overall interaction frequency of the recommendation data becomes more even, effectively improving the recommendation ecosystem and preventing the entire recommendation loop from falling into bias.

[0111] This embodiment also verifies the reliability of the bidirectional branch network principle and the effectiveness of the cumulative learning strategy through targeted ablation experiments. This method verifies the influence weight of each branch in the recommendation results by setting the dynamic adjustment parameters of the cumulative learning module to specific numbers such as [0, 0.2, 0, 5, 0.8, 1]. In addition to the unbiased test set used in the benchmark experiment, this method also randomly selected 10% of the samples from the original dataset as a test set with popularity bias. The results are as follows: as the parameter w increases, the model's effect on the biased test set gradually decreases, while its effect on the unbiased test set gradually increases, indicating that the model can gradually shift its attention to the long-tail side. However, when w = 0, the model consists entirely of long-tail branches, and the effect decreases, indicating that conventional learning branches have a significant promoting effect on modeling the basic characteristics of users and items representing the general public.

[0112] This embodiment effectively solves the problem of popularity bias in recommendation systems after application, avoiding the repeated recommendation of popular movies and the failure to discover long-tail items. At the same time, it can increase the diversity of recommendation results, avoid homogenization, and improve user satisfaction, demonstrating the usability of the present invention in removing popularity bias in recommendation systems.

[0113] The above description is merely one embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent recommendation based on a bi-branch graph neural network without popularity bias, characterized in that, The method includes: Obtain the original dataset; the original dataset contains several user item exchange records; Remove user item exchange records for popular items from the original dataset to obtain a long-tail dataset; Based on the original dataset and the long-tail dataset, an original subgraph and a long-tail subgraph are constructed respectively, wherein the nodes in the original subgraph and the long-tail subgraph include: user nodes and item nodes; For the original subgraph and the long-tail subgraph, information is propagated through graph neural networks. During the training process, the weight of the graph neural network corresponding to the long-tail subgraph is gradually increased based on a dynamically adjusted parameter, so as to obtain the user's item recommendation results based on the graph neural network corresponding to the trained long-tail subgraph. The step of constructing the original subgraph based on the original dataset includes: Set the first-order neighbors of the user node as item nodes with exchange records; Set the second-order neighbors of the user node to user nodes that have the same exchanged items; Set the third-order neighbors of the user node to the item nodes that have exchange records with the second-order neighbors; Set the third-order and above neighbors of the user node as high-order cooperative signals; The construction of a long-tail subgraph based on the long-tail dataset includes: Based on the user item exchange records in the long-tail dataset, an initial long-tail subgraph is constructed; wherein, the nodes in the initial long-tail subgraph include: user nodes and item nodes, the first-order neighbors of the user node are item nodes with exchange records, the second-order neighbors of the user node are user nodes with the same exchanged item, the third-order neighbors of the user node are item nodes with exchange records with the second-order neighbors, and the third-order and above neighbors of the user node are higher-order cooperative signals. The similarity between user nodes is calculated using the point-to-point mutual information method; Based on the similarity, edges are added between user nodes in the initial long-tailed subgraph to obtain the long-tailed subgraph.

2. The method as described in claim 1, characterized in that, After removing user item exchange records for popular items from the original dataset, a long-tail dataset is obtained, which includes: Based on the Pareto principle, user item exchange records for popular items are removed from the original dataset to obtain a long-tail dataset.

3. The method as described in claim 1, characterized in that, The step of adding edges between user nodes in the initial long-tail subgraph based on the similarity includes: Calculate users The set of items in the long-tail dataset Interaction probability of items ;in, Represents the set of items The number of items in the middle, Indicates user With the aforementioned collection of items The number of interactive items in the item; Calculate users The set of items in the long-tail dataset Interaction probability of items ;in, Indicates user With the aforementioned collection of items The number of interactive items in the item; Calculate users With users In the collection of items The probability of having the same interaction item. ; Based on the interaction probability The interaction probability and the probability To obtain the PMI value ; When the PMI value is greater than zero, in the case of the user With the user Weighted edges are added between the corresponding nodes, and the PMI value is used as the weight of the weighted edges.

4. The method as described in claim 1, characterized in that, The process involves propagating information through graph neural networks for both the original subgraph and the long-tail subgraph. During training, the weights of the graph neural network corresponding to the long-tail subgraph are gradually increased based on a dynamically adjusted parameter. This results in obtaining the user's item recommendation results based on the trained graph neural network corresponding to the long-tail subgraph, including: Obtain the true labels of the original data; The graph neural network corresponding to the original subgraph performs the first step. The information is disseminated in rounds, resulting in the first predicted label; The graph neural network corresponding to the long-tailed subgraph performs the first... The information is propagated in rounds, resulting in a second predicted label; Based on the first predicted label and the true label, calculate the first [predicted label] of the graph neural network corresponding to the original subgraph. Round loss function ; Based on the second predicted label and the true label, calculate the first [predicted label] of the graph neural network corresponding to the original subgraph. Round loss function ; Calculate the first The parameters of the wheel ;in, Indicates the total number of training rounds; Based on the Total loss function of the wheel Adjust the parameters of the graph neural network corresponding to the original subgraph and the graph neural network corresponding to the long-tailed subgraph; When the number of training rounds reaches the total number of training rounds. At that time, the graph neural network corresponding to the trained long-tailed subgraph is obtained; Based on the graph neural network corresponding to the trained long-tailed subgraph, the item recommendation results for the user are obtained.

5. The method as described in claim 4, characterized in that, The graph neural network corresponding to the original subgraph performs the first step... The information propagation in the round yields the first predicted label, including: For any node in the original subgraph vector representation By aggregating the vector representations of neighboring nodes, we obtain the vector representation. ;in, Indicates the number of aggregations. , The node represents the layer number of the graph neural network corresponding to the original subgraph. Represents a user node or an item node; According to the vector representation This yields the implicit representation of each node at each level. ; Generate a user-item interaction matrix based on users' historical interaction data. To obtain the adjacency matrix of the graph neural network corresponding to the original subgraph. ; Based on the implicit representation With the adjacency matrix Perform multi-layer forward propagation to obtain the implicit representation. ;in, Represents the angle matrix, This represents the nonlinear activation function of a graph neural network. Indicates weight; According to implicit representation Obtain the first predicted label.

6. A de-popularity biased intelligent recommendation device based on a dual-branch graph neural network, characterized in that, The device includes: The data acquisition module is used to acquire the raw dataset; the raw dataset contains several user item exchange records; The preprocessing module is used to remove user item exchange records of popular items from the original dataset to obtain a long-tail dataset; The graph construction module is used to construct an original subgraph and a long-tail subgraph based on the original dataset and the long-tail dataset, respectively, wherein the nodes in the original subgraph and the long-tail subgraph include user nodes and item nodes; The intelligent recommendation module is used to propagate information through graph neural networks for the original subgraph and the long-tail subgraph respectively. During the training process, the weight of the graph neural network corresponding to the long-tail subgraph is gradually increased based on a dynamically adjusted parameter, so as to obtain the user's item recommendation results based on the graph neural network corresponding to the trained long-tail subgraph. The step of constructing the original subgraph based on the original dataset includes: Set the first-order neighbors of the user node as item nodes with exchange records; Set the second-order neighbors of the user node to user nodes that have the same exchanged items; Set the third-order neighbors of the user node to the item nodes that have exchange records with the second-order neighbors; Set the third-order and above neighbors of the user node as high-order cooperative signals; The construction of a long-tail subgraph based on the long-tail dataset includes: Based on the user item exchange records in the long-tail dataset, an initial long-tail subgraph is constructed; wherein, the nodes in the initial long-tail subgraph include: user nodes and item nodes, the first-order neighbors of the user node are item nodes with exchange records, the second-order neighbors of the user node are user nodes with the same exchanged item, the third-order neighbors of the user node are item nodes with exchange records with the second-order neighbors, and the third-order and above neighbors of the user node are higher-order cooperative signals. The similarity between user nodes is calculated using the point-to-point mutual information method; Based on the similarity, edges are added between user nodes in the initial long-tailed subgraph to obtain the long-tailed subgraph.

7. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method as described in any one of claims 1-5.

8. A storage medium storing a computer program, wherein, The computer program is configured to execute any of the methods described in claims 1-5 at runtime.

Images