Research Description
Big data analytics in online social network for mental healthcare
Social network mental disorders (e.g., cyber-relationship addiction, net compulsion, information overload) have been noticed recently due to the emergence of online social networks. However, there is currently no system to detect such disorders using only the social network data. Therefore, we propose the first machine learning framework to identify potential patients using only the online social network data. We achieve 90% of accuracy, while the baseline which employs only the online duration achieves only 35%. Moreover, we also study and formulate the first therapy group formation problem to help these patients form therapy groups efficiently for receiving proper attention in time. We propose an approximation algorithm for the therapy group formation problem. These works appears in CIKM, WWW, IEEE Transactions on Knowledge and Data Enginerring (TKDE).
Graph mining and machine learning on graphs for finding suitable groups
In these works, we propose and study the graph mining and ML approaches about how to find groups in large social and spatial databases for different scenarios. The application scenarios span a wide spectrum, including 1) impromptu social activities, 2) selection of attendees along with the most suitable activity location, 3) forming quick response teams for disasters, and 4) selection of attendees to maximize friend-making likelihood. These works have been published in SIGKDD, VLDB, AAAI, IEEE TKDE, ACM TKDD, EDBT and PAKDD. They also received MOST Young Scholar Fellowship, MOST, Taiwan, and PAKDD Best Runner-Up Paper Award.
Deep Neural Network-related Research Topics
These works include graph-based recommendation systems, federated learning with blockchain, graph adversariay attacks and countermeasures, network compression, neural network watermarking.
Graph crawling, generation, and sampling with performance guarantees
In addition to group managements, we also propose the first graph generator that is able to preserve graph patterns and other important graph properties such as clustering coefficient and degree distribution. This generator is able to generate a billion-node graph within several minutes. This generator is published as free download. In addition, we also consider the problem of sampling multiple overlapped social networks with a statistical quality guarantee. The graph generator has appeared in ICDM, CIKM, ECML PKDD, TKDE, TBD.