Black Box Time Series Modeling

Keywords: time series modeling, machine learning, systems biology, surrogate models

Time series’ incorporate rich data and represent complex processes arising in a variety of applications, including systems biology, finance, environmental modeling, etc. The underlying processes that generate time series are often high fidelity and computationally expensive to evaluate. This makes time series analysis in such scenarios a time-consuming and slow endeavor.

This project will explore and study fast and efficient techniques to approximate computationally expensive time series datasets using machine learning techniques, with systems biology as a target domain. However, the approaches explored will be ‘black-box’ in nature, meaning that no problem-specific knowledge will be used in the modeling process. The solutions will be very general and applicable to time series problems from a variety of domains. Various machine learning model types such as neural networks, support vector machines, Gaussian processes, etc. will be evaluated. Optimal training methods will be developed that minimize model training time and retain model accuracy.

Tasks to be performed within the scope of the project include:

1. Literature survey of machine learning model types available for computationally expensive time series datasets.
2. Practically evaluating model types for time series encountered in systems biology.
3. Studying the robustness and expressive power model types for such time series.
4. Exploring model training methods that minimize training time while retaining model accuracy.
5. Formulating pointers towards black box modeling of time series in reaction networks in systems biology.

Prashant Singh <prashant.singh-at-it.uu.se>
Andreas Hellander <andreas.hellander-at-it.uu.se>