I must vote this paper :-). I like it. I asked for models...you produced one. Marcel Blattner [more]

Lessons to learn for the recommender system community: a) Testing new algos for recommender systems should be more than just 'delete half and predict the other half' of a given dataset. b) To design robust and sustainable systems, we have to elaborate on human decision mechanisms too. I am aware of the difficulties but I think the "low hanging fruits" in designing recommender systems are gone. To boost RecSys on the next level we have to incorporate different aspects than we have used so far. c) The community still lacks a flexible and robust benchmark framework. The way we test and compare our algorithms is too diverse. d) We have no 'models'. To have a model means to have a rough idea on how people take their decisions in a given setting. Again: low hanging fruits are gone. Marcel Blattner [more]

..nice paper. Just two remarks: a) In table 1 you give some 'characteristics' of the bipartite networks. Among others you report <k> and <d>. These numbers carry no information, since the investigated degree distributions are 'scale-free'. The mean is not a characteristic scale. b) It would be very interesting to compare C_{u} and C_{o} to the same quantities obtained from a random ensemble of networks with the same topological features - like degrees of users and objects. This would give a hint of the 'pattern relevance' you observe, compared to a purely random realization of the 'same' network. Regards Marcel Blattner [more]