Design of a dynamic and robust recommender system based on item context, trust, rating matrix and rating time using social networks analysis

Abstract

Collaborative filtering recommender systems type have been increasingly used in e-commerce sites both to facilitate users' decision-making and increase sales. On the one hand, the open and interactive nature of recommender systems makes them vulnerable to shilling attacks, and on the other hand, given that most recommender systems are used in dynamic environments, and this issue generates incremental data over time. One of the main obstacles of this type of recommender systems is the inability to model the dynamic behavior of users and the incremental flow of data, as well as the vulnerability to shilling attacks. Therefore, classic models do not have the necessary ability to provide suitable recommendations to the user and also to detect shilling attacks dynamically. The objective of this study was to address this gap by designing a dynamic and robust recommender system model against shilling attacks. This model was based on item context, trust, users' rating and users' rating time, and it benefits from the social networks analysis of users and items to suggest @N top items to the target user, and had a dynamic and increasing property over time and robust against shilling attacks. Finally, to assess the performance and robustness of the recommender system to shilling attacks, 4850 different tests have been performed without shilling attacks and under three average, random, and bandwagon attacks. To validate the proposed model, the results of the tests have been compared with similar methods such as TRACCF, TOTAR and T&TRS using the evaluation criteria of Precision, Recall, F1, MAE and RMSE.The results depicted that the proposed method, due to finding users and similar items in communities created by social networks, causes a reduction in the number of predicted items that are not liked by the user (FP) and the number of unpredicted items that are liked by the user (FN) and finally the F1 criterion (which is a combination of Precision and Recall criteria) performs better than the comparison methods. Also, this method is robust against the shilling attacks by detecting fake profiles and ignoring them in the recommendation process, and the evaluation criteria before and after shilling attacks show this.