Temporal point process (TPP) has been a decent mathematical framework for describing and modeling event sequences in the continuous time domain, which often carry additional attributes such as location, participants etc. Recently there are an increasing number of machine learning models for learning and inferencing with temporal point process, with a wide range of settings in understanding, predicting and intervening the dynamic behavior of different individuals, groups and systems. In the era of big data, effective learning with such temporal event sequences can be of important value to business and society, while traditional time series based methods often discretize the raw events into equal intervals, missing the continuous nature of timestamp.
In this tutorial, I will first give a basic introduction on the preliminaries of temporal point process, and revisit some popular and classic forms with a few application examples. The Bayesian form based learning will also be described. Then I will introduce recent advances in learning of point process, including deep learning and reinforcement learning for TPP. Finally I will show some new scenarios such as missing and censored observations for applying TPP models and discuss future directions, to facilitate further research in temporal point processes.
- Temporal point processes: Basics (15 minutes)
- Intensity function modeling
- Simulation of TPP
- Bayesian framework for TPP learning
- Typical models of Temporal point processes (30 minutes)
- Poisson process
- Hawkes process
- Self-correcting process (nonlinear Hawkes process)
- Time-varying Hawkes process
- Mixture models of Hawkes process
- Factorization-based Feature process
- MLE solution
- LS solution
- Deep learning for Temporal point processes (30 minutes)
- Neural temporal point process
- GAN for temporal point process
- Reinforcement learning for temporal point process
- Temporal point processes in practice (15 minutes)
- Learning from imperfect observations
- Inference from missing data
- Feature-based Random Stitching
- Super position
- Learning from warped sequences
- Applications: social network analysis, healthcare, recom- mendation, video trailer generation, crime analysis
- Open source toolboxes
Junchi Yan is a tenure-track independent research professor with Shanghai Jiao Tong University (SJTU), at both Department of Computer Science and Engineering and SJTU AI Institute. Before joining SJTU in April 2018, Junchi has been with IBM Research working on machine learning and computer vision research and applications since April 2011. During that time. he was once a Senior Research Staff Member and Chief Scientist for industrial inspection with IBM China Research Lab, and extensively applied temporal point pro- cess model for industrial preventative maintenance projects. He was also once a visiting researcher with IBM T.J. Watson Research Center (Yorktown Heights), Japan's National Institute of Informatics (Tokyo), and Tecent AI Lab (Shenzhen).
Liangda Li is a senior research scientist at Yahoo Research, in the search and search ads team. He lead the vertical search ranking, query understanding, search ads, query linguistic analysis projects of Science team. Before joining Yahoo Research, he earned his Ph.D. degree in the School of Computer Science at Georgia Institute of Technology, under the supervision of Professor Hongyuan Zha. He received his B.S. degree in ACM honored class, at the School of Computer Science from Shanghai Jiao-Tong University in 2010. He has been awarded the 2010 Microsoft Research Asia Young Research Fellow Award. His research interest includes machine learning and its applications in information retrieval and social network. In particular, his focus in on influence modeling in various real-world behavioral data, such as search intent understanding, urban intelligence and crisis/crime.