Título: Unsupervised Time Series Novelty Detection Using Clustering-based Local Autoencoders
Autores: Renan Fonteles Albuquerque, Guilherme de Alencar Barreto
Resumo: Novelty detection, also known as anomaly detection, plays a crucial role in identifying new or abnormal instances within a dataset. Traditional autoencoder models have been effective in learning compact representations of data, but they often struggle with capturing fine-grained local variations in complex and high-dimensional datasets. To address this limitation, we propose a novel learning method called Local Autoencoders (LAEs) for novelty detection. LAEs incorporate local information into the encoding and decoding processes, enabling more precise and detailed reconstructions. In this paper, we present preliminary results on evaluation of LAEs considering benchmark datasets for time series novelty detection and compare their performance against traditional (i.e., global) autoencoder and nearest neighbor learning method. The results demonstrate LAEs competitive performance in detecting novel instances, surpassing (in several cases) traditional autoencoders. The proposed LAEs present a promising avenue for further exploration in the field of novelty detection, leading into new opportunities for research and practical applications
Palavras-chave: Local learning, Autoencoders, Novelty detection, Time series
Páginas: 7
Código DOI: 10.21528/CBIC2023-172
Artigo em pdf: CBIC_2023_paper172.pdf
Arquivo BibTeX: CBIC_2023_172.bib