One of the keycharacteristics of Web is that it allows the Internet users to share theirviewpoints and opinions with other users about almost everything (Jakob, Weber,Müller, & Gurevych, 2009). Having another person’s substantiatedopinion can be of practical benefit when it comes to deciding whether or not toinvest time, money or effort into something. The ability of recommender systems to generate direct connections betweenusers and items that represent matches in interests and preferences makes theman important tool for alleviating information overload for users (Shi, Larson, & Hanjalic, 2014). Collaborative recommender systems aggregateratings or recommendations of objects, recognize similarities between users oritems on the basis of the ratings, and generate new recommendations (Schafer,Frankowski, Herlocker, & Sen, 2007; Su & Khoshgoftaar, 2009).Content-basedrecommendation is a continuation of information filtering research (Lops, DeGemmis, & Semeraro, 2011).
Content-based filtering recommendertechniques recommend items on the basis of the textual information of an item,under the assumption that users will like similar items to the ones they likedbefore (M. Pazzani& Billsus, 2007).Demographic recommendersystems categorize users or items based on their personal attribute and makerecommendation based on demographic categorizations (Son, 2014;Vozalis & Margaritis, 2004). Recommendersystems based on opinion mining are a more recent approach which employed textual reviews (Jakob et al., 2009; Musat, Liang, & Faltings, 2013). Each one of the above techniques uses different source of information toperform recommendation.Hybridrecommender systems combine two or more recommendation techniques to gainbetter performance with fewer of the drawbacks of any individual one (Burke, 2002;Lucas et al., 2013).
Theperformance of hybrid recommender systems can be improved by usingappropriately trained weight for each combined technique. However, the weightsof combined techniques may not equal for different users. This is due to thefact that some users have rated lots of items and less on other sources such ascomments and reviews and, therefore, they put more emphasis on rating than theother sources. On the other hand, some users leave lots of textual reviews andtheir tastes can be estimated based on these textual reviews. Therefore, forthese users, the weight of textual reviews should be more than the othersources of information. To consider these variations, we apply an individual K-dimensionalweight vector where K is the number of sources, for each user which determinesthe weight of each approach for that user. The proposed approach starts with aninitial user’s weights and, then, iteratively adjusts these weights so that theresulting weights maximize a suitable estimation of F-measure criteria.
Theadjustment rules are derived by solving the maximization problem throughgradient ascent.To summarize, we make several contributions. First, wepropose a general algorithm to handle different types of information on hybridrecommender systems and the algorithm is not limited to the used sources ofinformation. Second, we learn individual and personalized user weights for eachtechnique.
Third, we demonstrate that such a recommendation mechanism performssignificantly better than a standard one or than a recommender system based onless information, which it means adding extra information can help to recommendsmartly. Forth, wedirect extensive experiments on a film dataset to verify the usefulness of ouralgorithm.The rest of the paper is organized as follows. Insection 2 the background of this research is explained.
The proposed model isexplicated theoretically in section 3. Experimental results about theimplemented system along with a thorough analysis and a brief discussion aboutthe model is represented in section 4. Finally, section 5 concludes the paperand discusses the future work.