In this paper we evaluate the performance of different collaborative filtering algorithms over time, where new users, new items, and new ratings are constantly added to the recommender dataset.

The analysis has been performed on the datasets collected with ContentWise from two IPTV providers. Both datasets have been implicitly collected by analyzing the pay-per-viewmovies purchased by the users over a period of several months.

The first result of the paper outlines that item-based algorithms perform better with respect to SVD-based ones in the early stage of the cold-start problem. The second result shows that the accuracy of SVD-based algorithms, when using few latent factors, decreases with the time evolution of the dataset. On the contrary, SVD-based algorithms, when used with a large-enough number of latent features, increase their accuracy with time and may out perform the item-based algorithms if the dataset does not present a long-tail behavior.


PDF, 10 pages, 1.2 MB