Recommendation System Based on Competing Algorithms

In this paper, we provide the idea of analyzing the performance of algorithms generating personal recommendation by using competing algorithms for one and the same recommendation request based on same situation and information in a unified framework. The analysis of the recently proposed collaborative filter PAF (Popularity Among Friends) for finding user similarity based on past ratings and evaluation of missing ratings of a user-item-matrix in order to generate recommendations will serve as a base to specify competing algorithm for an experimental recommendation system. We present results of an on-line experiment of the proposed recommendation system which will demonstrate the advantage of directly comparing the rate of user acceptance of competing algorithms and allow a statement about their suitability as base of an easy evaluation of the system. The evaluation gives a conclusion about algorithms to be replaced through new competing methods in order to steadily improve the recommendation system.