Model factory : a new way to look at data through models

Abstract

In this thesis, we present Model Factory, a software framework that is able to generate predictive models from raw data. We present two foundational representations for data: an event-driven time series and a feature series. Together, they allow us to define a large suite of predictive modeling problems, and to subsequently solve them. We applied Model Factory to two real world datasets: one made up of sensor recordings from prototype cars, and the other containing time-varying status values for projects managed by a consulting firm. We deployed Model Factory on each of these datasets. Through the framework, we were able to enumerate a total of 3,877,848 predictive problems for the car dataset and 125,028 for the project dataset. We randomly sampled 150 and 1,000 prediction problems from the two datasets respectively, and solved them using off-the-shelf machine learning algorithms. We demonstrated our ability to build models for these prediction problems, and to gain insights into the data. We also built a graphical user interface on top of Model Factory for less tech-savvy users.