Time Series with PyTorch

17

Supervised and Unsupervised Anomaly Detection

In data analysis, the goal is not to find perfect patterns, but to identify when patterns break—for it is in the breaks that we find the signal.”—Andrew Ng

In previous chapters we built an understanding of temporal patterns and how they can be useful for identifying, segmenting, and modeling data. But what happens when the data we want to model contains observations that are unexpected or unlikely to be repeated? Detecting anomalies and outliers is a critical aspect of time series analysis, particularly when modeling with neural networks where unusual observations can distort learned representations.

Unlike classification, where we assigned data to known categories, or clustering, where we discovered ...