Unsupervised Deep Architectures for Attributed Network Embedding
Outliers can have an adverse effect on the embeddings of the other nodes in a network. This happens because the outlier nodes drive the random walks across communities and hence homophily property is violated in the resulting embeddings. To overcome this problem, the effect of outlier nodes in the overall embedding objective needs to be minimized while generating the embeddings of nodes. we proposed an unsupervised algorithm which reduces the effect of outliers in network embedding by using a matrix factorization approach. Further, to accommodate possible nonlinear representations, we proposed two unsupervised deep models.
Faculty: M. Narasimha Murty, CSA