Adaptive collaborative filtering for multi-objective Top-N recommendations with implicit feedback

Recommender systems (RSs) have emerged in several applications in response to the information overload challenge. Collaborative filtering (CF) achieved a remarkable prediction accuracy, but it has shown modest novelty, diversity, and coverage. Currently, most CF methods utilize explicit feedback of users (i.e., rating and reviews). However, most of the users’ behavior in online applications is considered implicit feedback (e.g., clicks, views, and purchases). Unlike explicit feedback, implicit feedback cannot directly represent user preferences. This study presents a new CF method for enhancing accuracy and personalized recommendations to users. The proposed method consists of a user profiling method, an adaptable interactions predictive model, and a Top-N ranking method. The user profiling method models the preferences of users. The interactions predictive model utilizes user profiles to customize the predictor to suit users’ context. The ranking method aims to select recommendations achieving relevance, novelty, diversity, and coverage objectives. This method involves an item filtering method to identify and exclude items with potential negative feedback from the recommendation list. A set of experiments was conducted to compare the proposed CF with alternative methods, such as matrix factorization and neural network-based methods. On average, the proposed CF achieved a 58.2% improvement in terms of prediction accuracy. Furthermore, the average achieved improvements in precision, novelty, coverage, and diversity are 37.9%, 738.95%, 40.8%, and 20.2%, respectively. Besides, the results reveal that the proposed RS mitigates the sparsity and cold start issues.