Time Weight Content-Based Extensions of Temporal Graphs for Personalized Recommendation

By Armel Jacques Nzekon Nzeko’o, Maurice Tchuente, Matthieu Latapy

In WEBIST 2017

Abstract

Recommender systems are an answer to information overload on the web. They filter and present to customer, a small subset of items that he is most likely to be interested in. Since user’s interests may change over time, accurately capturing these dynamics is important, though challenging. The Session-based Temporal Graph (STG) has been proposed by Xiang et al. to provide temporal recommendations by combining long- and short-term preferences. Later, Yu et al. have introduced an extension called Topic-STG, which takes into account topics extracted from tweets’ textual information. Recently, we pushed the idea further and proposed Content-based STG. However, in all these frameworks, the importance of links does not depend on their arrival time, which is a strong limitation: at any given time, purchases made last week should have a greater influence than purchases made a year ago. In this paper, we address this problem by proposing Time Weight Content-based STG, in which we assign a time-decreasing weight to edges. Using Time-Averaged Hit Ratio, we show that this approach outperforms all previous ones in real-world situations.

This entry was posted in Papers