Preferential Interpretation of Fuzzy Sets in E-shop Recommendation with Real Data Experiments

Our research aims to address the difficulties faced by small online retailers trying to provide meaningful recommendations to their customers. We hypothesize that fuzzy technology can be used to produce meaningful recommendations even in an environment with extremely sparse data. To set up our test, we found a small to mid-sized e-shop (SME) with no repetition (nR), an online travel agency. In other words, our test subject was an e-shop operating in a crowded market that it did not dominate selling a product where individual purchases were highly infrequent, often only annual. Its lack of dominance meant that it could not require pre-purchase registration, depriving it of one type of user-identifiable data. The purchase infrequency meant it could not assume cookie survival, depriving it of another source of user identifiable data. The only user identifiable data available was thus information mined using scripts or other time volatile techniques. To test our hypothesis, we first use fuzzy sets interpreted as user preference degree - preference indicators. Next we evaluated user-item data behavior using fuzzy preference indicators designed to build a rating for each new user. Next, we used content-based preference learning to expand this rating and so it could be used to infer preferences for yet unvisited items. Finally, we tuned the attribute preferences using the parametric families of t-conorms so that calculated user specific preference ordering could be extended to all available items. The quality of the solution was measured on top-k in several order sensitive metrics. To evaluate the effectiveness of our approach, we developed an off-line experimental framework that allowed us to test the results of our hypothesis against the actual data collected by the SMEnR´s web site. Based on our testing, we can conclude that a significant improvement in recommendation was achieved.