Popular online services provide a recommendation to users. Finding items fromthe recommendation list is accomplished by information retrieval using methodssuch as keyword search or browsing. In general, collaborative filtering algorithm isused to calculate of recommended items. Collaborative filtering utilizes a users ’history of purchases or rating scores for items. But there are some problems ofthis algorithm.1. A problem of presentation of recommendation items.Even if a user could see recommended items, he/she can ’t know the reason whythose items were recommended.2 A problem of reliance of recommendation items.Rating data of your non-similar user may influence on decision of recommendationitems for you.3. A problem of lacking the serendipity and novelty of recommended items.Users will expect to encounter some items which have serendipity or novelty whenusing recommender system. However, these will lack because recent collaborativefiltering techniques tend to use user’s profile too much and create detailed clusters.4. A problem of preparing enough data set of user’s ratings.To calculate recommend items with collaborative filtering, much amount of user’srating data is needed for the system.To resolve these problems, we propose these approaches below.At first, we developed a browsing system, named ZASH, for movie database. Thesystem used multiple 2D planes in 3D space so that the same type of information isdisplayed in each plane. The use of 3D space improved the visibility of links. Movietitles and commentators were laid out by using Multi-Dimensional Scaling methodso that the similar data are placed physically near each other. In this system,we considered additional information of the data as multi-dimensional data. Weexpected that this system can visualize relations between the data and additionaldata. Moreover, user would be able to see reliable recommendation items whichhave serendipity or novelty through similar user’s favorite items.Secondly, we analyze the differences in the results of the rating when the granularityof the score changes. In general, calculation of recommended items utilizes auser’s history of purchases or rating scores for items. We hypothesize that a binaryrating makes it easier for users to rate items and also makes it easier for recommendationsystems to perform calculations. This paper verified this hypothesisbased on the analysis of rating examples such as Lat.fm.
展开▼
