Marked Temporal Point Processes for Irregular Time Series in the Context of Medical Data and Recommendation Systems (2020)

by Kazi Tasnif Islam

Abstract: Many real-world systems consist of large volumes of event data measured at irregularly spaced time intervals. The time interval between two events carries much information about the system dynamics, alongside the type of the event and other associated data. One such example of a system is the intensive care unit (ICU), where the captured patient records consist of sparse, noisy, incomplete, heterogeneous, and unevenly sampled patients’ clinical data. Our first approach to model such data is a Bayesian structure learning based marked temporal point process model. We model the event streams, including vital signs and laboratory results in two different datasets using a piecewise-constant conditional intensity model (PCIM), a type of marked point process. Next, we model the stream of discrete clinical events in the continuous-time domain. We employ a neurally self-modulating marked temporal point process model that uses continuous-time long short-term memory (LSTM) cells as its building blocks. Our methods improve prediction performance in multiple tasks, including in-hospital mortality prediction, while providing suitable regularization and bypassing data imputation. We also experiment with a neural ordinary differential equation-based framework that simultaneously models the temporal point process’s intensity function and learn the underlying stochastic process that governs the distribution of the observed noisy measurements. This model achieves actual continuity as opposed to RNN based models, where the continuity is achieved through learnable decay rates of the hidden state, or through inputting the time between two observations as a scalar. Additionally, we apply a variational Bayes scheme to add uncertainty and regularization to the recurrent architectures on top of this framework. Another example of a real-world system containing a large amount of irregularly spaced time series-data is transaction systems, where the user’s buying behavior can be considered a highly irregular temporal sequence. Predicting the next return time of a user based on the transaction history is useful for providing the recommendations at the right time. In this regard, we develop a hierarchical session-based recommendation system that combines a discrete recurrent intra-session architecture and a continuous LSTM based inter-session model. We train the model to simultaneously predict a new session’s time using a temporal point process loss and recommend products within each session.

Download Information

Kazi Tasnif Islam (2020). Marked Temporal Point Processes for Irregular Time Series in the Context of Medical Data and Recommendation Systems. Doctoral dissertation, University of California at Riverside.

Bibtex citation

@phdthesis{Isl20,
   author = "Kazi Tasnif Islam",
   title = "Marked Temporal Point Processes for Irregular Time Series in the Context of Medical Data and Recommendation Systems",
   school = "University of California at Riverside",
   schoolabbr = "UC Riverside",
   year = 2020,
   month = Sep,
}

---

full list