Feature Selection for Time Series Classification & Adaptation for Ordinal Data
Feature Selection for Time Series Classification & Adaptation for Ordinal Data
Mor Abrutzky, Tel-Aviv University Advisors:Prof. Neta Rabin(TAU)
Abstract:
Feature selection is a common pre-processing task, which is typically aimed at reducing the dimension and the complexity of the dataset, while improving or maintaining model accuracy. While model performance is one important factor that is considered when performing feature selection, other factors like minimizing the number of features and incorporating the task constrains (like an ordinal target) are important and timely as data growth becomes a challenge. This work proposes two variants for a multi-class feature selection algorithm focusing on time series classification tasks, which have become more common over the last years. In the first part of the talk, I will review the main algorithmic approach. The separability strength of each feature with respect to all classes can be modeled by the Jeffries-Matusita (JM) matrix. The primary objective is to identify sparse feature subsets with complementary capabilities, preserving separability between each class pair. This process incorporates dimensionality reduction using diffusion maps (DM), which preserve pairwise distances from the high-dimensional feature space. Then, I will focus on feature selection for time series classification tasks. In particular, the work combines a recent time series representation method named ROCKET together with the feature selection algorithm. While ROCKET was shown to achieve state-of-the-art results for time series classification, it suffers from generation many features to represents the time series, this fact limits the classifiers that can be used. In addition, I will show how to adapt the method to the setting of feature selection for ordinary learning problems. The adaptation is performed by modifying the feature representation matrix. Last, I will propose a wrapper-type version of the feature selection method and demonstrate its performance on several examples while comparing it to known feature selection methods. Experimental results are demonstrated on various public datasets.
Bio:
Mor Abrutzky, M.Sc. student at the department of Industrial Engineering in Tel Aviv University, specializing in Data Science. Mor holds a B.SC. Degree in Industrial Engineering from Tel Aviv University. Her research, supervised by Prof. Neta Rabin.
