Adapting a recommendation system according to changing user consideration

The recommendation system is used to provide personalized recommendations to clients when they are selecting a product from a list of available options. The most often used technique for recommendation is collaborative filtering. In a collaborative filtering algorithm, the similarity factor employed in discovering the users with the same actions is one of the critical components of recommendations based on user rating for items; it does not consider any other information related to the recommender system entity (user and item). Although these methods can produce successful predictions, they have some drawbacks, such as scalability issues, finding suitable neighbors, and facing the cold-starting problem. Thus, the similarity computed in this manner is volatile for user pairs sharing a limited set of experiences and is undecidable for most users due to the absence of specified common tastes. In this dissertation, different solutions are presented to identify fluctuating user preferences and provide accurate recommendations. The first solution proposes a novel neighbor selection approach based on correlation and slope that focuses on determining the relevance of entities. It is outlined the connected entities' similarities according to importance degree. Additionally, we proposed a new model for rating prediction based on the first accessible rating from nearby neighbors. The second solution introduces a new user similarity measure to promote recommendation accuracy by determining each user's similarities even when only a few ratings are available. The suggested model considers only users' common ratings based on the proportion of dissimilar ratings and the highest value of this dissimilarity. The third method provides a user similarity algorithm that uses item genre characteristics to provide more accurate suggestions. It establishes a relationship between users based on item genre preferences, identifies the current users' nearest neighbors in each genre, and generates a final list of nearest neighbors who share the same taste across all genres. Finally, it uses a definite prediction process to forecast ratings for goods. Additional to suggest a solution to the user cold-start problem via the "Ask-to-rate" technique/adaptive approach of inquiring about the new user's chosen item genre. Then, top users from the same genre are considered nearest neighbors for the active user. To assess the model's accuracy, we suggest two additional assessment criteria: the rating level and the rating level's reliability among users. Numerous tests on real-world data sets have been undertaken. The experimental results demonstrate that the suggested algorithms generate highly accurate and reliable predictions.