Chow, Foo and Manai (all at Group Digital Life, Singtel, Singapore) 2014.
PDF.
The authors consider a personalised PageRank on a graph whose vertices represent items for recommendation and whose edge weights are determined by feature overlap (where features are e.g. category, tag). In order to incorporate collaborative information into the computation, they define the “teleport vector” (or “reset vector”) for a user to be the sum over all items they’ve interacted with of the corresponding rows of the behavioural item x item matrix (they call these “user correlation matrices”, somewhat confusingly).
This is a nice advance from “ItemRank” for incorporating item meta information into the recommendation process. In contrast to ItemRank, the authors work in the context of implicit feedback data. However, I think that the approach could be made more elegant by considering the users and tags (for example) as additional vertices in the graph – the reset vector would then just be the one-hot vector singling out the vertex corresponding to the user receiving the recommendations.
The authors carry out an enormous user survey and an impressive live production test to demonstrate the performance of their approach.