Top-N recommender system and effect of shilling attack onmulti-criteria rating values

Recommendation systems have popular ways of helping users generate estimates for a product and create product lists based on users' feedback. The outputs of such systems can increase customer satisfaction by generating accurate recommendations. However, the success of the systems depends on the level of customization of the user preference data. Therefore, it is possible to increase this level of personalization by collecting more detailed customer data. These systems, called multi-criteria recommendation systems, provide more accurate recommendations to the user with these increased levels of customization. However, when we look at the studies in the literature, it is seen that there are no top-n recommendation systems developed on multi-criteria systems. In this thesis, firstly, a new neighborhood creation process and an intuitive fuzzy set-based top-n recommendation system method are proposed for multi-criteria datasets. The proposed method consists of two important points: i) determining the relational structure between products; ii) the investigation of user trends, distinctive structures, and rating distributions. While the relational structure between the rating distribution and products is determined by the association rule and Entropy measurement, the attitudes and tendencies of the users are analyzed using intuitionistic fuzzy sets during the evaluation. The performance comparison of the proposed approaches is carried out by adapting the algorithm in the single-criteria systems to the multi-criteria system. In addition to user satisfaction, the proposed list is based on different measures such as serendipity, diversity, and novelty. It has been seen that the suggested methods offer very successful/correct recommendations to the user, and user satisfaction is achieved. In addition to providing successful recommendations to the user, the vulnerability to shilling attacks adversely affects the performance of recommendation systems. Therefore, in the second stage of the study, an analysis of how robust the item-based multi-criteria top-n recommendation systems are against shilling attacks is carried out. In the study, four different variants of the power item attack type, which is a very successful new attack model in item-based systems, are used. An experimental design including new and existing evaluation criteria for the effectiveness of attacks, an attack model adapted to multi-criteria systems, a target, and strong product selection is used. According to the experimental results based on the real-world dataset, it has been seen that the multi-criteria top-n algorithms are open to manipulations. While various methods exist to defend these manipulated systems against such attacks, obtaining robust systems by capturing shilling profiles remains difficult for those based on multi-criteria ratings. Therefore, in the last phase, a classification method is proposed that detects fake profiles based on user characteristics. In addition to the general and model-based methods available in the literature, new features based on product popularity and user characteristics have been added to these user features. From the experimental results, it is concluded that the proposed method successfully distinguishes attack profiles from real users even when the selected size and attack size are small. Empirical results also show that product popularity and user attributes based on rating profiles are highly useful features in capturing shilling attack profiles.