Representation learning, aiming to automatically learn low-dimensional latent features for use in various scientific research projects and business applications, is an important research topic in the fields of data mining, machine learning and big data analytics. In this project, the research team developed new representation learning frameworks with novel ideas to capture rich, meaningful and discriminative features in various kinds of heterogeneous information networks, including social networks, bibliographic networks, citation networks, road networks, etc., in order to support a variety of network data analysis and mining applications, such as restaurant type classification, patent classification, travel time estimation, future citation prediction, route planning, and so on. This project investigated generic network representation learning techniques by exploring multiple relationships specified in forms of meta-paths in a heterogeneous information network and developed efficient and effective neural network models to learn latent representation of nodes and meta-paths in the network. While generic network representation learning techniques can be generally applied to a wide range of heterogeneous information networks, representation learning techniques for specific types of heterogeneous information networks, e.g., road networks and publication citation networks where the moving behaviors of travelers on roads and the flow of knowledge among published papers are very different, need to be specialized to explore inherent characteristics and unique features of the corresponding networks and application domains. Specifically, this project investigated representation learning problems on road networks and publication citation networks and developed new representation learning frameworks with novel ideas, neural network model designs, data preparation and sampling schemes.  With rigorous testing and comprehensive evaluation using datasets collected in the project, the developed frameworks models are shown to outperform the state-of-the-art techniques by extensive experiments, which validates the various new ideas proposed in the project. While these algorithms, models and techniques are designed for the learning frameworks in this project, the underlying ideas and concepts resulted from the research may very much likely be applicable to other neural network architectures, potentially advancing the research in data mining and machine learning. 

This project has generated data, publications, presentations, and software that may facilitate follow-up research and collaborations. The research findings have been disseminated to the research communities of data mining, machine learning and data science through publication and conference presentations. The research insights obtained from the project have further inspired the research team to explore new research problems in knowledge graphs, transfer learning, reinforcement learning, and community detection. This project allowed the team to achieve its educational objectives by teaching and promoting student learning in the fields of data mining and machine learning. Two Ph.D. students involved in the project have graduated with solid training in data mining and machine learning. The project also facilitated the team to provide research opportunities to M.S., undergraduate and female students. Several undergraduate students (including a female) involved in the project entered graduate schools after graduation.  

Last Modified: 12/14/2021
Modified by: Wang-Chien Lee