Skip to Main content Skip to Navigation

Metric learning for structured data

Abstract : Metric learning is a branch of machine learning, and more specifically representation learning, which involves learning the parameters of a metric in order to improve its intrinsic qualities when using it. In this thesis, we summarize recent developments around metric learning, both on so-called flat data, but also on structured data. The lack of power of representation of the metrics currently used leads us to propose a new metric based on sub-modular functions. This new proposal significantly increases the evaluation of interactions between data dimensions. A learning algorithm for this metric is then proposed. Secondly, we present two approaches LSCS (selection of relational constraints) and MRML (multi-relation metric learning) allowing to process types of data extremely present in information systems: relational databases. All of these proposals are then compared experimentally with existing ones, justifying the practical interest of the proposed approaches.
Complete list of metadata
Contributor : Hoel Le Capitaine Connect in order to contact the contributor
Submitted on : Thursday, March 25, 2021 - 7:13:55 PM
Last modification on : Wednesday, April 27, 2022 - 3:48:27 AM
Long-term archiving on: : Saturday, June 26, 2021 - 6:07:13 PM


Files produced by the author(s)


  • HAL Id : tel-03180284, version 1


Jiajun Pan. Metric learning for structured data. Artificial Intelligence [cs.AI]. Université de Nantes, 2019. English. ⟨tel-03180284⟩



Record views


Files downloads