Papers^γ

# In short Short article considering HIN-based recommendation. The contribution consists in using not only meta-paths in the HIN, but also "enhanced" meta-paths which are related to more elaborate motifs (typically triangles) # Summary ### 1. Introduction - usually HIN-based RS rely on the number pf meta-paths from u to i, the more there are, the higher the recommendation - Figure 1 summarizes situations that the authors want to distinguish: (u1,b4) and (u1,b3) are equivalent in terms of (P2) meta-paths, but not in terms of 3-nodes based "trust" patterns - patterns here are considered using 3 nodes patterns (à la Milo et al) as represented in Figure 2, meta-paths based on these motifs are called Motif Enhanced Meta Path (MEMP) - the concept is explored on 2 datasets: Epinions and CiaoDVD ### 2. Framework - to compute similarity using MP, we build the adjacency matrix W_{Ai,Aj} where Ai and Aj are types, then we build the commuting matrix Cp which is a product of the adjacency matrices corresponding to path P => similarities are based on counting obtained by matrix products ##### MoHINRec framework - they define adj matrix based on the same principle, but an edge is considered if it belongs to a pattern of interest (illustrated on Figure 3) - equation 2: they merge different matrices (obtained with usual MPs or MEMPs) with an alpha-weighting - then they test this with state-of-the-art HIN-based RS (actually, one they have designed in another work) ### 3. Experiments and analysis - 2 datasets, described in Table 1: - Epinions: ~22K users, 300K items, 900K ratings - CiaoDVD: ~17K users, 16K items, 72K ratings - evaluation metrics: MAE, RMSE (accuracy based) - baselines for comparison: RegSVD (MF with regularization), SoReg (MF using social links for regularization), SocialMF (MF with social trust propagation), FMG (IN-based RS, state-of-the-art) - supposedly better than the others - Settings: 8/1/1 train-validation-test, 5 experiment series with different splits ##### Performance comparison: - Summary in Table 2: MoHINRec outperforms the others - depending on the motif considered, we should vary the alpha coefficient to obtain the best possible performance (it is not clear for me how they tune alpha) - performances vary depending on the dataset and morif considered - FMG is still more efficient than othher benchmark methods ### 4. Related work ##### 4.1 HIN based recommendation ##### 4.2 motifs (very complex networks based: Milo et al., recent works by Leskovec et al.