Drifting Preferences in Recommender Systems

Abstract

Recommender systems are increasingly becoming popular with the enormous choice that the online virtual marts present. Collaborative filtering is one of the most popular techniques to generate recommendation by means of collaboration among multiple information agents. It uses past transactions to gather critical information and then extracts knowledge by means of filtering. One of the major issues in collaborative filtering is the sparsity problem, wherein the data is sparse in nature and carries only partial information or misses out information totally. Another issue is that in reality, collaborative filtering is characterized by the recency effect wherein recent items tend to speak volumes about user preferences than past data. This concept, sometimes called the drift effect is absent in the traditional collaborative filtering algorithm. In this paper, an attempt has been made to come up with a novel approach that would try to address the sparsity problem and would take the drifting effect into consideration. This algorithm uses minimal information to make predictions and takes the drifting effect fully into consideration. Some newer algorithms do make use of a decreasing time function that assigns a maximal weight to the recent data and a minimal weight to past data. However, if the time-frame from which the data is constructed is not continuous in nature, this approach has the possibility of assigning differential weights to a data cluster that belongs to the same time-frame. Here, a sliding window approach has been taken that views cluster of data in consecutive frames of time in a window. The algorithm applies a decaying or decreasing function only when a particular trait falls outside the window by means of consecutive incidence of non-occurrences. The user-preference is clustered under five distinct heads, namely “Sporadic”, “New”, “Regular”, “Old” and “Past”. Initial occurrence places a particular trait into “Sporadic” category. Repeated occurrence of this particular trait to fill a window of defined size promotes this trait to the “New” category. Repeated occurrence of this particular trait to satisfy a bigger reference window of given size promotes this particular trait into the “Regular” category. Conversely, repeated non-occurrences of the particular trait in the “New” category to void a particular window of pre-determined size would demote the trait to the “Sporadic” category. Further non-occurrence of the particular trait would pull itself out of our reference heads completely. Conversely, repeated occurrences of the particular trait would promote it to the “New” category once again and subsequently back to the “Regular” category. Once a trait is in the “Regular” category, several repeated non-occurrence would put it into an intermediate category called “Past” that denotes that the interest has started fading off and then moves to “Old” and then moves away. The algorithm has been tested on a sample of Yahoo! Movies community’s preferences for various movies