Phuong, Thang and Phuong (all from the Posts and Telecommunications Institute of Technology, Vietnam), 2008.
PDF obtainable here.
The authors present a recommendation system that incorporates user ratings of items (they work in the explicit rating context) and item-feature relations. The approach is graphical. Effectively, two weighted graphs with non-negative weights are constructed, network propagation is performed on both independently and the two resulting scorings are combined in a weighted sum. The first graph is directed represents user-item preference via the item features (where the user-feature preferences are computed heuristically from the given ratings and item-feature associations); thus in this graph all paths from user to item have length 2. The second graph is undirected and represents the purely positive user ratings of items and excludes the item features. The two graphs capture the content- and collaborative- aspects of the recommendation problem, respectively.
I find the approach lacks unity and is too heuristic. The unity suggested by the user/item/feature graph of Figure 1 is merely pictorial, since network propagation is actually performed on the two graphs described above (which are derived from this unified graph) separately. The two separate graphs are constructed heuristically, and this removes any claim the approach might have had to necessity.
The more obvious, unified, approach (network propagation on the user/item/feature graph) is unavailable here since the user-item associations may be negative. This would not be the case if the feedback were implicit (e.g. purchases). For this reason I will be interested to read the paper of Huang et al. cited by the authors – perhaps they use just such an approach. Furthermore, Huang et al. experiment with different network propagation algorithms (in this paper, a modification of an algorithm by Weston et al. is used).
The MovieLens dataset is used in the evaluations, which demonstrate the superiority of the authors approach over a purely content, a purely collaborative and a simple hybrid approach that merges the result sets of collaborative filtering and content recommendation computed separately.
Paper is clearly written.