Collaborative Filtering Recommendation-Based Random Negative Sampling and Graph Attention

Weiqiang Li

doi:10.1109/ACCESS.2025.3541126

Collaborative Filtering Recommendation-Based Random Negative Sampling and Graph Attention

2025-01-01

0

OA

AI

DOI：10.1109/ACCESS.2025.3541126

原文链接

原文求助

摘要

En 中文

In this era of information overload, recommendation systems significantly enhance the efficiency of information delivery, better meeting the needs of users. Currently, GCN-based recommendation systems typically use degree normalization or mean pooling to aggregate neighbor messages. These methods learn embedding representations for users and items. However, both message-passing mechanisms overlook the varying importance of different neighbor nodes to the target node. As a result, the learned representations of users and items are not sufficiently accurate. Furthermore, they do not take into account the importance of data augmentation of metadata, which limits the recommendation performance. On the other hand, during the model's loss optimization process, sample imbalance is prone to occur. Negative samples greatly outnumber positive samples, which leads to a certain degree of overfitting in the model. To address this issue, this paper proposes a collaborative filtering recommendation based on random negative sampling and graph attention (NGACF). First, perform data augmentation on the initial embeddings of users and items. Then, before the propagation of each layer in the embedding propagation layers, graph attention networks (GAT) are used to aggregate information from the target node's neighbors. This approach captures the importance of different neighboring nodes, thereby enriching the target node representations. In the loss optimization module, a random negative sampling strategy is incorporated as an auxiliary loss to mitigate the problem of imbalanced sample classes during training. This approach reduces model overfitting, facilitating better optimization and improving the model's generalization ability. Finally, experiments were conducted on three public datasets. In particular, on the Amazon-Book dataset, the results show that the proposed method outperforms the baseline model. Recall@20 improved by 3.65%, and NDCG@20 increased by 2.86%. These results further validate the effectiveness of the proposed model.

Keyword:

Collaborative filtering

Vectors

Training

Recommender systems

Computational modeling

Optimization

Collaboration

Attention mechanisms

Overfitting

Metadata

Recommendation system

collaborative filtering

graph attention

data augmentation

random negative sampling

Collaborative Filtering Recommendation-Based Random Negative Sampling and Graph Attention

摘要

期刊

机构

引用论文