On novelty detection for multi-class classification using non-linear metric learning

Author(s)

Silva, Samuel Rocha ^[1] ; Vieira, Thales ^[2] ; Martinez, Dimas ^[3] ; Paiva, Afonso ^[1]

Affiliation

^[1] Univ Sao Paulo, ICMC, Sao Carlos - Brazil

^[2] Univ Fed Alagoas, Inst Comp, Maceio, Alagoas - Brazil

^[3] Univ Fed Amazonas, Dept Matemat, Manaus, Amazonas - Brazil

• Document type: Journal article
• Source: EXPERT SYSTEMS WITH APPLICATIONS; v. 167, APR 1 2021.
• Web of Science Citations: 0

Abstract

Novelty detection is a binary task aimed at identifying whether a test sample is novel or unusual compared to a previously observed training set. A typical approach is to consider distance as a criterion to detect such novelties. However, most previous work does not focus on finding an optimum distance for each particular problem. In this paper, we propose to detect novelties by exploiting non-linear distances learned from multi-class training data. For this purpose, we adopt a kernelization technique jointly with the Large Margin Nearest Neighbor (LMNN) metric learning algorithm. The optimum distance tries to keep each known class' instances together while pushing instances from different known classes to remain reasonably distant. We propose a variant of the K-Nearest Neighbors (KNN) classifier that employs the learned distance to detect novelties. Besides, we use the learned distance to perform multi-class classification. We show quantitative and qualitative experiments conducted on synthetic and real data sets, revealing that the learned metrics are effective in improving novelty detection compared to other metrics. Our method also outperforms previous work regularly used for novelty detection. (AU)

• FAPESP's process: 19/23215-9 - Meshfree methods based on generalized finite differences using MLS and SPH
• Grantee: Afonso Paiva Neto
• Support Opportunities: Regular Research Grants