Feature Hashing And Multi-task Learning Based Network Representation Learning

Network is an important way to present the relationships between objects,such as social network,State Grid and citation network.With the increasing complexity of network,there is more valuable to explore that as a carrier of information.There are some meaningful applications in network analysis,such as node classification,link prediction,community detection and recommender system.However,there is a limitation when machine learning applied into network analysis due to its high dimensionality and sparsity.Thus,it is a promising research how to properly construct the structure information extracted from networks into meaningful presentation.Network representation learning,also called network embedding,is proposed to encode network information into a continuous low dimensionality feature space.From network topology,those nodes who have similar structure should have similar representation vectors.For example,those nodes within in a same community have similar proximity structure and thus they should be closer in embedding space.Due to the learned representations,the relationships between nodes and roles that nodes play in networks could be efficiently analysis.Not only that,those representations can be deeply applied into other application tasks.Thus,network representation learning is a promising research.We proposed tow methods for network representation learning.Our main contributions are as follows.(1)Network representation leaning based on feature hashing.The algorithm firstly extracts high order proximity structure through random walks and build proximity matrix.Then feature hashing is introduced to reduce the dimensionality of this matrix and thus nodes embeddings are generated.For solving the duplicate colliding during the reduction of dimensionality,a new approach feature hashing with multi kernels is proposed,resulting in decreasing colliding probability.And parameter sensitivity experimental results confirm the effectiveness of our proposed multi-kernels feature hashing.(2)Network representation learning based on multi-task learning.Traditional network representation learning approaches only focus on how to extract the high order proximity structure but ignore the nodes one-hop area structures.In order to more meaningfully character the complex relationships between nodes,we proposed a network representation learning approach based on multi-task learning,which both preserve the global and local feature.There are two tasks in our model.The first is to preserve the global feature by extracting the high order proximity structure from pointwise mutual information matrix.The second is to preserve the local feature by reconstruction of nodes' one hop area.