Metric-Guided Multi-task Learning

Jinfu Ren, Yang Liu, Jiming Liu*

*Corresponding author for this work

Research output: Chapter in book/report/conference proceedingConference proceedingpeer-review


Multi-task learning (MTL) aims to solve multiple related learning tasks simultaneously so that the useful information in one specific task can be utilized by other tasks in order to improve the learning performance of all tasks. Many representative MTL methods have been proposed to characterize the relationship between different learning tasks. However, the existing methods have not explicitly quantified the distance or similarity of different tasks, which is actually of great importance in modeling the task relation for MTL. In this paper, we propose a novel method called Metric-guided MTL which explicitly measures the task distance using a metric learning strategy. Specifically, we measure the distance between different tasks using their projection parameters and learn a distance metric accordingly, so that the similar tasks are close to each other while the uncorrelated tasks are faraway from each other, in terms of the learned distance metric. With a metric-guided regularizer incorporated in the proposed objective function, we open a new way to explore the related information among tasks. The proposed method can be efficiently solved via an alternative method. Experiments on both synthetic and real-world benchmark datasets demonstrate the superiority of the proposed method over existing MTL methods in terms of prediction accuracy.

Original languageEnglish
Title of host publicationFoundations of Intelligent Systems
Subtitle of host publication25th International Symposium, ISMIS 2020, Graz, Austria, September 23–25, 2020, Proceedings
EditorsDenis Helic, Martin Stettinger, Alexander Felfernig, Gerhard Leitner, Zbigniew W. Ras
Place of PublicationCham
Number of pages11
ISBN (Electronic)9783030594916
ISBN (Print)9783030594909
Publication statusPublished - 17 Sept 2020
Event25th International Symposium on Methodologies for Intelligent Systems, ISMIS 2020 - Graz, Austria
Duration: 23 Sept 202025 Sept 2020

Publication series

NameLecture Notes in Computer Science
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349
NameISMIS: International Symposium on Methodologies for Intelligent Systems


Conference25th International Symposium on Methodologies for Intelligent Systems, ISMIS 2020

Scopus Subject Areas

  • Theoretical Computer Science
  • Computer Science(all)

User-Defined Keywords

  • Metric learning
  • Metric-guided multi-task learning
  • Multi-task learning
  • Task relation


Dive into the research topics of 'Metric-Guided Multi-task Learning'. Together they form a unique fingerprint.

Cite this