MEMN:Multiple Vectors Embedding for Multi-Label Networks

Network embedding, which assigns vectors to network nodes in a manner that preserves the network features, is a hotspot of network research in recent years. A salient common feature of the existing approaches is that each node is mapped to exactly one vector. This one-vector mapping is insufficient to represent the nodes’ attribution in those extensively existed networks whose nodes’ have multiple labels. In this paper, we present MEMN, a novel approach of multiple vectors embedding for multi-labeled networks. For any node in the network, MEMN employs Node2vecWalk to generate its neighbor nodes. We maintain a neighbor cluster center for each label of the node and induce its label by clustering the embeddings of the neighbor nodes. Then, we assign vectors, one per label, to the node. This method can be non-parameterized, namely, NP-MEMN method. That is, if the number of label vectors for a node is not given, NP-MEMN can learn during embedding. Empirical studies on real datasets show that either MEMN or NP-MEMN outperforms many widely used methods in both multi-label classification and link prediction.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.