In today's Internet era, the vast amount of information and the challenges of resource selection by users require proper guidance. Recommender systems have emerged as vital tools to assist users in selecting resources and obtaining information. These systems are designed to generate intelligent recommendations for specific groups of users, especially those interested in particular products. In recent decades, various methods for recommender systems have emerged, and diverse algorithms, articles, and scientific texts have been introduced in this field. Graph neural network techniques have gained popularity as an effective solution for modeling the graph structure of recommender systems. These techniques provide better representations for users and items by collecting information from neighbors. In this study, graph neural network techniques are used to construct a continuous-time bipartite graph to represent all temporal interactions. This graph includes vertices representing users and items, and edges with temporal features for each interaction. These temporal features assist in sorting item purchase sequences and modeling temporal dependencies of interactions, distinguishing the temporal effects of neighbors. By defining a graph convolution operation based on position and time, temporal patterns and dynamics between user-item interactions are modeled. This method learns dynamic representations for users and items simultaneously in the bipartite graph and models complex high-order connections between users and items using multiple layers of graph convolution. In recommendation datasets, there are usually a limited number of popular items with high repetition, which receive a lot of attention. In contrast, rare items with low repetition are less present. This difference causes the recommendation model to focus only on popular items. To solve this problem, a diverse recommendation approach using re-weighting methods for highly repeated items has been introduced. In this approach, different weights are assigned to different categories of items to make the model pay attention to recommending rare items as well. Additionally, recommender methods usually focus on removing false negative samples. In this study, instead of removing these samples, a method is used to detect false negative samples. This method transforms these samples into positive samples and uses them to improve the model performance. Finally, the MovieLens dataset is used to eva‎luate this method, and the results show that the proposed method outperforms other baseline and deep methods. The article continues to discuss the basic concepts, related works, details of the proposed method, implementation, and obtained results.

عبدالرضا ميرزايي

مهران صفاياني

سمانه حسيني , شيرين بقولي زاده