Applying collaborative filtering to reputation domain: a model for more precise reputation estimates in case of changing behavior by rated participants

Abstract

Automated Collaborative Filtering (CF) techniques have been successfully applied on Recommendation domains. Dellarocas [1] proposes their use on reputation domains to provide more reliable and personalized reputation estimates. Despite being solved by recommendation field researches (e.g. significance weighting [2]), the problem of selecting low-trusted neighborhoods finds new roots in the reputation domain, mostly related to different behavior by the evaluated participants. It can turn evaluators with similar tastes into distant ones, contributing to poor reputation rates. A Reputation Model is proposed to minimize those problems. It uses CF techniques adjusted with the following improvements: 1) information of evaluators taste profiles is added to the user evaluation history; 2) transformations are applied on user evaluation history based on the similarities between the taste profiles of the active user and of the other evaluators to identify more reliable neighborhoods. An experiment is implemented through a simulated electronic marketplace where buyers choose sellers based on reputation estimates generated by the proposed reputation model and by a model that uses traditional CF. The goal is to compare the proposed model performance with the traditional one through comparative analysis of the data that is created. The results are explained at the end of the paper.