Hybrid approach for recommendation system

Abstract

The primary objective of recommendation systems (RSs) is to analyze user’s fondness and taste and recommend similar items to him/her. There exist various methods, e.g., user/item collaborative filtering (CF), content-based filtering (CBF), association rule mining (ARM), hybrid recommender system (HRS), for recommendations. Though these methods possess excellent characteristics, they are inefficient in providing good recommendations in particular situations. For example, item CF produces recommendations for cold-start objects; however, it typically has low accuracy compared to user CF. Conversely, user CF often provides more accurate recommendations; however, it fails to provide recommendations for cold-start objects. The hybrid methods aim to combine different approaches coherently to yield better recommendations. This paper presents an HRS based on user CF, item CF, and adaptive ARM. The proposed HRS employs ARM as a fundamental method; however, it considers only a set of users who are nearest to the target user to generate association rules (ARs) among items. Also, the support levels to mine associations among items are adaptive to the number of rules generated. Results of the study indicate that the proposed HRS provides more personalized and practical suggestions compared to the traditional methods.