CAREER: Towards A New Paradigm of Automated, Robust, and Adaptive Feature Engineering

In real-world scenarios, data structure is usually noisy, redundant, and complex. If real-world data is input into a machine learning model directly, it may lead to inefficient training and, ultimately, poor performance. Features are used to describe the characteristics of data. Feature engineering is a fundamental preprocessing step, which optimizes raw data's feature space so that it can be used by the model more effectively. However, traditional feature engineering approaches are usually manual, time-consuming, and difficult to generalize across different domains. To deal with these problems, this project will develop new technologies to make feature engineering automated, robust, and adaptive. The research outcome will benefit a variety of critical applications, where feature engineering plays a crucial role, including disease diagnosis, vehicular security, emergency medical service, software engineering, material reliability, and electronic commerce. The developed algorithm prototypes have good potential to be adopted by technology companies to improve their real-world products and services. This project supports the development of new curricula, undergraduate student research, internships, and local K-12 education. This project will introduce an operator-centric feature engineering paradigm, which unifies feature generation, feature transformation, and feature selection through a multi-agent reinforcement learning framework. Specifically, (1) to address framework challenges, the project will first propose a novel learning framework by designing reinforcement learning policies to drive operator agents. The action space of operator agents is optimized by a compression strategy. Agent collaboration is coordinated by a multi-level information sharing and interaction mechanism. Computational efficiency is improved by a high-efficiency cache and incremental update mechanism, as well as a parallelization mechanism; (2) to address data challenges, the project will first tackle noise by introducing an environmental perturbation mechanism during training. A sparse activation strategy is designed to deal with high-dimensional data. An adaptive feature alignment and knowledge transfer mechanism is proposed to handle heterogeneous data, and a two-stage technique is proposed to support fusion of multi-modal features; (3) to address task challenges, the project will first design a dynamic decomposition and joint optimization mechanism to improve generalizability across multiple tasks. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2543905 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT,01003031DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Kunpeng Liu | Institution: Clemson University, CLEMSON, SC | Award Amount: $368,275 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2543905 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2543905.html