Ranked criteria profile for multi-criteria rating recommender

Abstract

Recommendation process consists of user profile creation, neighbor formation and prediction from neighbors' opinion. If user profile can closely represent user's characteristic, high quality neighbors will be formed and accurate results will be obtained consecutively. Recently researchers are interested in multi-criteria user profile to represent user's preference on multiple aspects. Most of them are created as a vector of the preference valued on each criterion. The current ranked criteria profile is created based on idea that highly preferred criterion will get high rank order. However, preference level of criterion may be opposite to overall score which indicates whether user will select an item. In this paper, the significance level of each criterion affecting to overall score is discovered and integrated into ranked criteria user profile. Moreover, either ROC rank weighting technique or score mapping table is applied to compare a pair of ranked profiles containing rank value which is an abstract number and could not be comparable. The experimental results show that incorporating a new ranked criteria user profile and score mapping table encourages getting better results than current multi-criteria rating recommendation methods. © 2012 Springer-Verlag.