Building and Evaluating an Adaptive Real-time Recommender System

Most recommender algorithms in use today are slow to adapt to changes in user preferences. This is because they are focused towards model-building and offline calculation of recommendations. The fact that they require large amounts of information about users before they can make sensible recommendations does not help their case either. This work proposed an adaptive prediction scheme that makes real-time recommendations to users. The scheme was developed by Kristiaan Pelckmans[1]. It is real-time in that it calculates new recommendations every time a user submits some side information. It is adaptive in that it maintains an online memory of user activities which evolves as user preferences change. In this work, the current start-of-the-art in the implementation of recommender systems is investigated. The adaptive prediction scheme is explained in detail. Its applicability in driving a recommender system is evaluated in comparison with other “established” recommender algorithms. Using a movie recommender system implemented using the scheme, it is shown that the scheme relies on much less data in order to make recommendations and the quality of its recommendations is slightly better than the common recommender algorithms which are based on collaborative filtering. Lastly, the scheme’s limitations are highlighted and recommendations for future work are made