Invited Speaker

Georgiana Ifrim:
Effective Linear Models for Learning with Sequences and Time Series

Abstract: In this talk I present some of the work done in my research group on developing machine learning algorithms for classification and regression tasks on sequences and time  series data. The focus is on algorithms to train linear models. We show that albeit these linear models are considered too simple to achieve high accuracy in many learning tasks, when trained in rich feature spaces they are strong competitors to very complex models such as ensembles and deep learning models. Linear models with rich features are as accurate as complex non-linear models, but are very efficient to train and are interpretable. Interpretability in this context means that the model (a list of weighted features) and the prediction (a sum of feature weights) is transparent to the user. I first provide an overview of important and wide application areas where we  encounter sequences and time series, discuss common approaches to learn with sequences, and present algorithms for sequence classification and regression tasks. I also show how ideas from sequence learning can naturally carry over to time series data and show that a linear model with features selected from multiple symbolic representations, achieves state-of-the-art time series classification accuracy. By combining multiple representations of the sequence data to create rich features, we enable linear models to achieve high accuracy, have efficient training and preserve interpretability, the latter being a crucial requirement in many applications.

Slides: here

Web site: