The effects of demographic information in matrix factorization recommender systems on accuracy and serendipity for varying levels of user feedback

In this thesis we investigate the influence of demographic information in a matrix factorization recommender systems on the accuracy and serendipity of its recommendations. In particular we investigate possible disparate effects for different levels of user feedback available in addition to the global effects on all users. In line with previous research we have found that adding demographics on average improves accuracy. The performance advantage does seem to diminish as more feedback is available per user, yet there is no evidence to belief that the demographics will ever become noise to the model. In terms of serendipity the consequences of demographic information are twofold. On one hand they lead to less novel recommendations on average. On the other hand demographics lead to recommendations with higher unexpectedness given the user’s context. The performance differences are however small and again seem to diminish for higher levels of feedback available per user. In general it is therefore beneficial to include demographics in a matrix factorization recommender as it alleviates the cold-start problem, while doing little harm to serendipity. Only when the ultimate goal is selling the most novel items it would be advised to omit the user’s demographic background for the recommendation.