Re-evaluation of Recommendion Systems by Serendipity metric

Abstract

By inputting a Keyword its related Information spreading on Web can be easily collected. The users often read the text by skim reading in front of the vast information. But we think that it may be possible that the skipped information is useful, which we cannot find only by developing higher level Recommender Systems. In these years we concider not only the Recommender Systems accuracy but also the evaluations and proposals by Serendipity metrics have been studied. There have been many metrics proposed by the researchers, but we don't have any standardized one yet. In this research, by using two metrics, the already-developed Recommender System, we examine if it recommended Serendipity information to the users: one is Yuan’s serendipity metric and another is Adamopolous' serendipity metric. Yuan's metric showed a high value 0.9055, instead Adamopolous’ metric showed a low value 0.1262. Adamopolous metric seeks the quantity included in all recommended lists. The Serendipity information does not appear frequently, so we expect the values are inevitably low. Accordingly I think we should create an metric by considering and understanding that the Serendipity information is not the information which often appears.