Venue: Westin Houston, Memorial City, Houston, TX, U.S.

Date: April 20th, 2024

Tutorial Description

Feature selection has invariably been a fundamental step in supervised classification towards enhancing a machine learning model's accuracy. In many real-world scenarios (e.g., medicine and criminal justice), a feature is acquired at a cost (e.g., medical tests, evidence collection), which can relate to, among others, the time and effort spent in generating it, and its discriminative power. At the same time, accurate, time-sensitive, and interpretable classification is needed. For instance, in clinical diagnosis, informative tests (e.g., magnetic resonance imaging) may be intrusive but informative.

In this tutorial we will systematically review the state-of-the-art in cost-sensitive feature selection. The goal is three-fold:

• clarify the notion of cost-sensitive feature selection, and related nomenclature,
• summarize progress on this growing research area,
• draw connections to recent trends in machine learning (e.g., model interpretability)
• identify challenges and opportunities for future work.

Tutorial Outline

• Introduction [5 mins]
• A simple machine learning problem
• Definition of feature selection and variants
• Applications and main challenges
• Online/Streaming feature selection [40 mins]
• Problem definition
• Online/streaming feature selection
• Variants (e.g., streaming data, feature interactions, group feature selection)
• Instance-wise feature selection [40 mins]
• Problem definition
• Static approaches
• Dynamic methods
• Advanced Topics [30 mins]
• Model interpretability
• Incorporating fairness constraints
• Dealing with structure (e.g., multidimensional Bayesian network classification, and hierarchical classification)
• Conclusion [5 mins]

Benchmark datasets as well as real-world applications including cyberbullying detection, creditworthiness prediction, streamlining delivery of physical services, and computational medical diagnosis will be used for illustration purposes.

The tutorial slides can be found here: part 1, part 2 and part 3.

Presenters

Dr. Daphney-Stavroula Zois is Associate Professor in Electrical and Computer Engineering at the University at Albany. She specializes in machine learning and statistical signal processing with a particular focus on decision making under uncertainty. Before joining UAlbany, she was a Postdoctoral Researcher at the University of Illinois at Urbana Champaign. She obtained her Ph.D. from the University of Southern California in 2014. Dr. Zois serves as a reviewer of high-impact journals such as IEEE TSP, IEEE TAI, IEEE TNNLS, and conferences such as AAAI, NeurIPS, ICLR and ICASSP.

Dr. Charalampos Chelmis is Associate Professor in Computer Science at the University at Albany, where he leads the Intelligent Big Data Analytics, Applications, and Systems group. He specializes in socially important data science. He is acting Associate Editor of SNAM journal, senior program committee member at AAAI, and program committee member of international conferences, including WSDM and the WebConf. He is also reviewer for journals including IEEE TKDE and IEEE TCSS.

Both presenters are actively working on cost-sensitive feature selection and have made several contributions towrads advancing the field. Their work has been published mainly at the IEEE Transactions on Artificial Intelligence (TAI), and the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).