Hybrid link prediction via model averaging

It is a big challenge to select and combine multiple similarity-based approaches in a proper way for link prediction. Recently, some studies have focused on hybrid link prediction, but neglected to select models for link prediction. Actually, model averaging is more promising to reduce the risk of misspecified. Therefore, this paper proposes a novel linear model to integrate various kinds of local indices, and employs two typical model averaging approaches, Smooth Akaike Information Criterion (S-AIC) and Smooth Bayesian Information Criterion (S-BIC), for hybrid link prediction. It is worth noting that the objective function motivated by KL divergence matches our model quite well without unconfirmed links, which indicates node pairs of observed links should have larger similarities than unconnected node pairs. Experimental results on six datasets demonstrate that the proposed method can achieve more accurate performance. Our work provides a promising way for hybrid link prediction, and is a preliminary exploration to study on model averaging for high dimensional matrices.