Papers^γ

# Short summary - The authors propose a Recommender System which is a hybrid between collaborative filtering and content-based recommendation - Content-based recommendation is essentially based on a HIN structure, with various kinds of content nodes (Figure 2 shows an explicit example) - In the broad lines, their approach consist in making a random walk on the HIN structure - More precisely, this is a RW with restart (which allows to personalize the results), and this is a Vertex Reinforced Random Walk, i.e. a specific kind of RW where future transitions to a node are more probable if this node has been visited in the past - VRRW are not Markovian processes, and this translates into the fact that the transition matrix should be updated - There is a lot of technical details about the implementation, which could be useful to someone who want to apply the method practically - The learning process is achieved through Stochastic Gradient Descent - They evaluate the efficiency of their approach (Div-HeteRec) on two Meetup datasets: one in Bangalore, another in Hyderabad (general stats in Tab.1) - Performances are compared to RWR, CF-NMF and two versions of their method with less parameters (Uni-HeteRec, without personalization ; Learn-HeteRec, with a static transition matrix) - Results in terms of accuracy (precision, recall and NDCG @k, with k=1,2,3,5,10) are compiled in Tables 2 to 5 for two different reco problems: group-to-user and tag-to-group - Div-HeteRec performs better for group-to-user recommendation (perf better on Hyderabad dataset than on Bangalore dataset) - Learn-HeteRec performs better for tag-to-group recommendation - They provide hypothetical justifications to their observations