A performance analysis on time-series-based recommender system

Abstract

With the ever-growing information technology, information overload has become an important issue faced by today’s society. Recommender system, as a way of dealing this problem, has become an important tool in people’s daily life and has been widely adopted in different businesses. In academia, it has also become a popular topic that keeps attracting the attention of researchers from many research fields. Among the papers that are published on this topic, most of the methods focus on personalized recommendation. Those methods usually adopt the idea of collaborative or content-based filtering and view the interactions between user and item as a matrix format. Their evaluation is also based on offline evaluation which usually does not consider the interactions’ global timeline. This project, on the other hand, focuses on non-personalized recommender system. Instead of considering the interactions between user and item as a matrix, the algorithm proposed in this project only cares about how many times an item is interacted. In other words, only the item’s popularity is taken into account when making recommendations. In this report, a non-personalized recommender, which is based on the idea of considering item’s popularity over time as a time series, is proposed. The training and evaluation of this recommender and the baseline methods also follow the global timeline and are performed in a manner that is similar to prequential evaluation. Moreover, by this evaluation and some additional explorations on the data itself, this project also provides an analysis on the temporal characteristics of the data and the recommender’s performance on the data.