We solve critical and challenging business and societal problems by designing novel and rigorous machine learning algorithms and methods, rather than applying existing machine learning algorithms. A typical research project starts with an important business/societal problem, identifies its characteristics that cannot be effectively catered by existing machine learning algorithms, and designs a novel machine learning algorithm/method to solve the problem by addressing these characteristics effectively and efficiently. In our research, we strive to make methodological contributions to both problem domains and machine learning. This stream of research is termed as Use-Inspired AI research by the National Science Foundation.
We are interested in designing state-of-the-art analytical tools to solve fundamental problems of Financial Technology (FinTech). We actively explore new opportunities enabled by recent development in machine learning to provide better solutions than traditional methods. One project aims at developing a better way of assigning firms to their corresponding industries by designing a novel deep learning algorithm to analyze firms’ annual 10K reports. Another project develops algorithms that can learn hidden relationships among training instances (e.g., hidden relationships among loan applicants) and leverages these learned relationships for better predictions (e.g., predicting the probability of default).
- Xiaohang Zhao, Xiao Fang, Jing He, Lihua Huang. Exploiting Expert Knowledge for Assigning Firms to Industries: A Novel Deep Learning Method. Accepted by MIS Quarterly.
- Hongzhe Zhang, Wei Qian, Xiao Fang. Latent Network Information-Enhanced Credit Risk Prediction.
Social Network Analytics
We focus on a unique phenomenon in social networks – the diffusion of contents, information, and adoption behaviors from one social entity to another. In particular, we investigate three critical and related problems concerning this phenomenon: adoption, persuasion, and link recommendation. That is, the diffusion of contents, information, and adoption behavior is initiated by persuaders and reached to adopters through the linkage structure of a social network. Accordingly, we study the following problems: how to predict adoption probabilities in a social network? how to predict top persuaders in a social network? and how to recommend links for a social network?
- Fang, X., and Hu, P. 2018. Top Persuader Prediction for Social Networks. MIS Quarterly 42 (1), pp. 63-82.
- Li, Z.*, Fang, X.*, Bai, X. and O. R. Liu Sheng. 2017. Utility-based Link Recommendation for Online Social Networks. Management Science, 63(6), pp. 1938-1952. (*Equal contribution).
- Fang X., Hu, P., Li., Z., W. Tsai. (2013). Predicting Adoption Probabilities in Social Networks. Information Systems Research. 24(1), pp.128-145.
We investigate fundamental data science problems. For real world applications, new data are continuously generated. Newly generated data could bring in new knowledge and invalidate part or even all of the earlier discovered knowledge. As a result, a fundamental problem in data science is how to maintain the currency of knowledge discovered from rapidly evolving data sources, namely the knowledge refreshing problem. Data incompleteness is another common problem in data science. For example, product adoption decision in a social network context depends on social influence, entity similarity, structural equivalence, and hidden factors. While it is easy to collect data for the first three factors, hidden factors are unobserved, i.e., observed data are incomplete. Therefore, it is necessary to study how to discover knowledge from incomplete data.
- Fang X., Liu Sheng, O. R., P. Goes. (2013). When Is the Right Time to Refresh Knowledge Discovered from Data? Operations Research.61(1), pp. 32-44.
- Fang X., Hu, P., Li., Z., W. Tsai. (2013). Predicting Adoption Probabilities in Social Networks.Information Systems Research. 24(1), pp.128-145.