DESIGN OF WEB PERSONALIZATION TECHNIQUES USING SOFT COMPUTING

Abstract

The advancement of personalized recommendation systems is essential in to day’s digital landscape, where user-specific content and services play a piv otal role in enhancing the user experience. However, these systems face significant challenges, particularly in balancing the trade-off between ex ploration and exploitation of user preferences and ensuring explainability in decision-making algorithms. This thesis addresses these challenges by proposing several novel solutions, namely FuzzyBandit, Hybrid-Neuro Ban dit, and Contextual-POI-Bandit, aimed at improving web personalization. The FuzzyBandit model optimizes decision-making through a dynamic feed back mechanism. This model adjusts parameters based on the relevance and diversity of features to maximize rewards while maintaining interpretability by generating explanations for its decisions. A novel trust score framework is also developed, assessing the reliability of the model’s outputs, thus enhanc ing user trust and enabling the detection of potential errors in the system’s reasoning. Further advancing the field, this research presents the Hybrid Neuro Bandit (HNB) model, which integrates the most effective expert advice from exist ing recommendation models while discarding those that underperform. This approach demonstrates advantages across various datasets and scenarios, ad dressing the variability in model performance. Additionally, the Contextual POI-Bandit framework is introduced, which integrates Social, Geographical, Temporal, and Categorical (SGTC) contextual influences into a unified user vector. This framework significantly enhances the predictive accuracy of per sonalized Point of Interest (POI) recommendations by closely analyzing user behavior and anticipating future visits. Furthermore, all the solutions have been empirically evaluated on benchmark datasets, comparing them with state-of-the-art models across performance metrics such as recall, precision, and accuracy, offering substantial improve ments in the effectiveness, reliability, and interpretability of the developed personalized recommendation models.