CTNRL: A Novel Network Representation Learning With Three Feature Integrations

Related Works

The early work of network representation learning is mainly based on matrix eigenvector calculation. This includes local linear representation and Laplace feature table. This method requires high spatial complexity and time complexity, which makes such algorithms unable to be used on large networks.

Inspired by the Word2vec (Mikolov, Sutskever, et al., 2013; Mikolov, Chen et al., 2013; Mikolov et al., 2015) algorithm, the DeepWalk algorithm introduces a neural network into network representation learning. The algorithm first carries out a random walk on the network. Then, it inputs the node sequence into the neural network to obtain the vector representation of nodes. The subsequent Node2vec (Grover & Leskovec, 2016) algorithm and LINE (Tang et al., 2015) algorithm are inspired by DeepWalk algorithm. Unlike methods like DeepWalk, which uses shallow neural networks, the SDNE (Wang et al., 2016) algorithm uses an unsupervised deep autoencoder for the training of network nodes. Different from the DeepWalk algorithm and SDNE algorithm based on the near neighbor hypothesis, the struc2vec (Ribeiro et al., 2017) algorithm believes that two nodes that are not close neighbors may have high similarities.