Extracting Homogeneity and Heterogeneity of Different Learning Tasks: A Novel Multi-Task Learning Framework for Feature Extraction, Regression, and Spatio-Temporal Prediction

This project aims to develop and demonstrate a novel multi-task learning (MTL) framework to address the fundamental challenges of extracting the homogeneity and heterogeneity of different learning tasks for effective multi-task feature extraction, regression, and spatio-temporal prediction.

MTL solves multiple related learning tasks simultaneously and thus improves the learning performance of all tasks. It has attracted extensive research interest and has become particularly important in recent years. Nevertheless, MTL remains to be fundamentally challenging due to three difficult research questions: (i) How can the intrinsic homogeneity/commonality of different tasks be captured and its physical meaning be made interpretable? (ii) How can the heterogeneity/individuality of different tasks be characterized to improve the learning performance of each task without sacrificing other tasks' performance? (iii) How can the various MTL models be integrated into a unified framework that is adaptable to different learning problems/scenarios?

To answer these questions, we propose to develop a general and flexible MTL framework, making full use of the information from all tasks to discover and integrate the homogeneity and heterogeneity of the given tasks. Using the developed framework, we will further propose three new MTL methods for various learning problems:
(1) MTL-F for feature extraction, which will learn the compact and interpretable representations of the original data for all tasks, with the homogeneity and heterogeneity of different tasks being well preserved;
(2) MTL-R for regression, which will extract the homogeneity and heterogeneity to directly enhance the effectiveness and efficiency of multiple learning tasks; and
(3) MTL-ST for spatio-temporal prediction, which will capture complex dependencies of spatio-temporal data in real-world situations and provide accurate predictions.

We will (a) formally define and formulate the learning problems; (b) theoretically analyze the algorithmic behavior of the developed framework and methods; (c) empirically validate solutions on both synthetic and real-world datasets; and (d) practically implement and evaluate our framework and methods to address a challenging real-world problem: malaria risk prediction in the Greater Mekong Subregion and Southeast Asian countries.

The developed framework and methods will contribute to the development of artificial intelligence and machine learning in academia. More importantly, the project will accelerate the elimination of malaria in GMS and SEA. This is critical for achieving the ultimate target of globally eradicating this life-threatening infectious disease. The project will also create a plenty of research and practice opportunities for both undergraduate and graduate students so as to nurture the local talents in the field.