Novelty Detection in Time Series via Weak Innovations Representation: A Deep Learning Approach

TL;DR

This paper introduces a deep learning method for novelty detection in time series by causally extracting innovations that are statistically independent of past data, enabling online detection of structural changes.

Contribution

It presents a novel deep learning framework that causally extracts innovations for online novelty detection with proven minimax optimality and robustness.

Findings

Effective detection of novel changes in real and synthetic datasets

Proven minimax optimality under Bayes risk measure

Robustness demonstrated through experiments

Abstract

We consider novelty detection in time series with unknown and nonparametric probability structures. A deep learning approach is proposed to causally extract an innovations sequence consisting of novelty samples statistically independent of all past samples of the time series. A novelty detection algorithm is developed for the online detection of novel changes in the probability structure in the innovations sequence. A minimax optimality under a Bayes risk measure is established for the proposed novelty detection method, and its robustness and efficacy are demonstrated in experiments using real and synthetic datasets.