Article Preview
Top1. Introduction
Recommendation is an expressed opinion regarding the quality of an entity (item, activity, etc.). Recommendations assist us in making decisions when we do not have sufficient personal experience in a specific area of interest. The development of recommendation approaches is highly affected by the advancement of different media channels. The following forms of recommendation are presented chronologically:
- •
Word of Mouth: This early form of recommending is passed on by one individual to another and usually involves trusted friends or family members.
- •
Expert Reviews: These reviews can be tied to the printed press; for example, critique columns for movies and restaurants or places of interest. One of the earliest examples of book reviews is found in “The New York Times Book Review”, a weekly review magazine supplement starting from 1896.
- •
Content Based Recommendation: Content based recommendations are derived from the field of information retrieval and based on the similarity of items. This recommendation form can be tied to early electronic applications, designed for the scientific community. Early examples of this form for scientific paper recommendation are reviewed by Housman, & Kaskela, (1970).
- •
Collaborative Filtering (CF): CF uses similar features between users (user stereotypes/user behavior) to recommend items based on other users' behavior. The term “Collaborative Filtering” was coined by Goldberg, Nichols, Oki, & Terry, (1992). Their work described a mailing system (tapestry) that uses knowledge gained from other users to filter incoming messages for a specific user. In this system, users collaborated by recording their reactions to documents they have read. These reactions were later used to implement new filtering rules.
- •
Recommender Systems (RS): RS is an inclusive term utilizing all or some of the above methods to create a recommendation for a user. RS are used to suggest data items to people who are likely to be interested in them. The term “Recommender systems” was introduced by Resnick, & Varian, (1997). RS, following the categorization given by Belkin, & Croft, (1992) is an application of information filtering.
The task of evaluating RS is widely addressed in the literature. The importance of the field is so high, that it has emerged as an independent research field by itself (Mulwa, Lawless, O'keeffe, Sharp, & Wade, 2012), encouraging scientists to collaborate and identify better and clearer evaluation methodologies (Amatriain, Castells, De Vries, & Posse, 2012). Previous works have discussed the evaluation of RS and addressed the problem of developing efficient metrics to measure the effectiveness of recommendations (Adomavicius, & Tuzhilin, 2005; Mulwa, Lawless, Sharp, & Wade, 2011). Evaluating RS, is usually based on some metrics, which evaluate the recommendation success. Since RS differ in their output to the user, a broad range of metrics have been used in the literature. Some works have offered to include various dimensions of evaluation criteria, mainly (a) Serendipity – the generation of unexpected and novel recommendations (Mcnee, Riedl, & Konstan, 2006), (b) Coverage – the degree to which recommendations cover the set of available items and the degree to which recommendations can be generated to all potential users (Ge, Delgado-Battenfeld, & Jannach, 2010), (c) Utility – avoidance of bad and trivial recommendations (Meyer, Fessant, Erot, & Gaussier, 2012) and (d) Stability – the consistency of RS predictions (Adomavicius, & Zhang, 2012). However, the lack of datasets and the need for consolidation and comparison of recommendation methods cause scientists to revert to the classical evaluation metrics (Amatriain et al. 2012).