Amirali Changizi
Effectiveness of Masked Autoencoder Pre-Training for Supervised Time-Series Anomaly Classification: A Controlled Study.
Rel. Santa Di Cataldo, Alessio Mascolini, Francesco Ponzio. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (6MB) | Preview |
Abstract
Time-series anomaly detection is critical in domains such as healthcare, industrial monitoring, and finance, yet labeled anomalies are scarce and expensive. Recent time-series foundation models suggest that self-supervised pre-training on large heterogeneous corpora may reduce this dependence on labeled anomaly data, but the specific contribution of pre-training is often difficult to separate from architectural capacity, adaptation strategy, and benchmark design. This thesis studies whether Masked Autoencoder (MAE) pre-training on a large heterogeneous corpus (LOTSA) improves downstream supervised anomaly classification on the UCR Anomaly Archive. The work is designed as a controlled study: under a matched protocol (same encoder architecture, classifier head, data splits, windowing, optimization budget, and evaluation rules), we compare three adaptation regimes.
In every regime the downstream classifier is the same trainable multilayer-perceptron (MLP) head; the regimes differ only in how the encoder beneath it is treated: (i) a LOTSA-pretrained encoder kept frozen, (ii) a LOTSA-pretrained encoder that is partially fine-tuned, and (iii) an architecturally identical encoder trained from random initialization on the target data only
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
