An Approach with Low Redundancy to Network Feature Selection Based on Multiple Order Proximity.

Most models for unsupervised network feature selection use first-order proximity and reconstruction loss together as a guiding principle in the selection process. However, the first-order proximity is very sparse and insufficient in most cases. Moreover, redundant features, which can significantly hamper the performance of many machine learning algorithms, have seldom been taken into account. To address these issues, we propose an unsupervised network feature selection model called Multiple order proximity and feature Diversity guiding network Feature Selection model (MDFS), which uses multiple order proximity and feature diversity to guide the selection process. We use multi-order proximities based on the random walk model to capture linkage information between nodes. Moreover, we use an auto-encoder to capture the content information of nodes. As a last step, we design a redundancy loss to alleviate selecting highly-overlapping features. Experiment results on two real-world network datasets show the competitive ability of our model to select high-quality features among state-of-the-art models.