Broad Structural Representation Learning

A broad spectrum of data from different sources and structures are widely existing, such as natural graph data (social network, IT/OT network, brain network), unnatural graph data (image, text, sphere), sequence data (stock). Modeling these data with heterogeneous sources and structures is a fundamental problem in data mining with diversified applications in many science and business fields. Given the intrinsic heterogeneous nature, broad visions and strategies for structural representation are required to derive competitive advantages and unlock the power of the big data. We investigate and develop novel deep learning approaches for structural pattern analysis and discovering in the graph. Specifically, we proposed new representation learning models from the graph data via graph neural networks. The graph data provides a generalized representation of many different types of inter-connected data collected from various disciplines. Besides the unique attributes possessed by individual nodes, the extensive connections among the nodes can convey very complicated yet important information. The graph data are very challenging to deal with because of their complex structures (containing multiple kinds of nodes and extensive connections), and diverse attributes (attached to the nodes and links). To address these issues, I will show how to develop structural preserving and heterogeneity preserving representation learning model taking the benefit of graph neural network. I also apply the board structural learning on multiple applications including, healthcare, cybersecurity, recommender system, and natural language processing.