Joint Similarity Learning for Predicting Links in Networks with Multiple-type Links

This paper addresses the problem of link prediction on large multi-link networks by proposing two joint similarity learning architectures on nodes’ attributes. The first model is a similarity metric that consists of two parts: a general part, which is shared between all link types, and a specific part, which learns the similarity for each link type specifically. The second model consists of two layers: the first one, which is shared between all link types, embeds the objects of the network into a new space, while the second one learns the similarity between objects for each link type in this new space. The similarity metrics are optimized using a large-margin optimization criterion in which connected objects should be closer than non-connected ones by a certain margin. A stochastic training algorithm is proposed, which makes the training applicable to large networks with highdimensional feature spaces. The models are tested on link prediction for two data sets with two types of links each: TED talks and Amazon products. The experiments show that jointly modeling of the links given our frameworks improve link prediction performance significantly for each link type. The improvement is particularly higher when there are fewer links available from one link type in the network. Moreover, we show that transfer learning from one link type to another one is possible using the above frameworks.