M.A. Thesis:

An automated trust-based recommendation method using regression adaptive neuro-fuzzy inference system (ANFIS) in e-commerce applications

Researchers consider the recommender system as a subset of support decision-making systems that can be used in various processes such as counseling, purchasing, and giving proper advices. Despite the fact that in the previous solutions a variety of models have been proposed, they suffer from shortcomings such as lack of proper ground for building user trust, insecurity in the systems and software bugs with regard to the delivery of proper advices to users. In addition, the impossibility of learning for sample data, failure to address the issue of uncertainty and abundant errors in responses cannot be overlooked. In this thesis, an efficient strategy and a hybrid model of machine learning and image processing techniques have been proposed that can be used to create an advisory system. This research is implemented in three phases: (1) using the Walker trust model for the user; (2) using the graphics encryption algorithm to provide security for the user; and (3) constructing the support decision-making unit as regression and classification to provide a proper advice to the user based on neural network trained by the regression adaptive fuzzy neural network algorithm with genetic evolutionary algorithm. The results of implementation both in trust phase and creation of a safe environment for the user and in assessing the performance of the recommender decision-making system were desirable. In the trust-building phase, factors such as coverage, precision and F Measure were tested and an error of less than 8% was achieved in the phase of creating safe advisory environment. For the regression decision maker, minimum error, maximum error, MAPE, and MSE were considered in offering proper advice. Finally, classification precision, performance and uncertainty in the conventional classification of advisory system labels was taken into account. The values analyzed in this thesis are at desirable level and their values in the multi-class mode, training stages and experiments varied in the range of 88% to 90%.

Keywords: Recommender system, trust ground, security, neural network, regression adaptive fuzzy neural network