In general, a practical online recommendation system does not rely on only one algorithm but adopts different types of algorithms to predict user preferences. Although most of similarity measures can rapidly calculate the similarity on the basis of co-rated items, their prediction accuracy is not satisfactory in the case of sparse datasets. Making full use of all the rating information can effectively improve the recommendation quality, but it reduces the system efficiency because all the ratings need to be calculated. To recommend items for target users rapidly and accurately, this paper designs a hybrid item similarity model that achieves a trade-off between prediction accuracy and efficiency by combining the advantages of the two above-mentioned methods. First, we introduce an adjusted Google similarity to rapidly and precisely calculate the item similarity in the condition of enough co-rated items. Subsequently, an intuitionistic fuzzy set (IFS) based Kullback-Leibler (KL) similarity is presented from the perspective of user preference probability to effectively compute the item similarity in the condition of rare co-rated items. Finally, the two proposed schemes are integrated by an adjusted variable to comprehensively evaluate the similarity values when the number of co-rated items lies in a certain range of value. The proposed model is implemented and tested on some benchmark datasets with different thresholds of co-rated items. The experimental results indication that the proposed system has a favorable efficiency and guarantees the quality of recommendations.

Introduction

The rapid growth of the Internet has tremendously boosted online enterprises, especially e-commerce, providing consumers with a wide variety of choices in books (Amazon), videos (YouTube) and photos (Flickr) (Baluja, Seth, Sivakumar, Jing, Yagnik, Kumar, Ravichandran, & Aly, 2008; Brynjolfsson, Hu, & Smith, 2003; Zheng, Li, Liao, & Zhang, 2010), etc. However, the massive amount of information on the Internet usually overwhelms users and makes them indecisive. (Liu, Hu, Mian, Tian, & Zhu, 2014). Recommender systems (RS) have been successfully deployed to provide information, make recommendations, and facilitate decision-making on products of interest for active users (Davidson, Livingston, Sampath, Liebald, Liu, & Nandy, 2010; Shahabi, Banaei-Kashani, Chen, & Mcleod, 2001; Takeuchi, & Sugimoto, 2006). They can match users’ expectations and points of interest by analyzing their previous preference behaviors of users, thereby addressing the information overload problem effectively. As one of the best-known recommendation techniques, collaborative filtering (CF) (Breese, Heckerman, & Kadie, 2013) has been adopted by numerous e-commerce websites. It provides unknown items to the target users by learning the potential interests of the users. The general recommendation process of CF involves three main steps. The first step calculates the similarity degree among users. The second step selects the most similar users with the target users as the nearest neighbors. Finally, the third step predicts the preferences of users and recommends items for them.