Anomaly Detection from Time Series Under Uncertainty

Anomalies in data can cause potential issues in downstream tasks, making their detection critical. Data collection processes for continuous data are often defective and imprecise. For example, sensors are resource-constrained devices, raising questions about their reliability. This imprecision in measurements can be characterized as noise. In machine learning, noise is referred to as data (aleatoric) uncertainty. Additionally, the model itself introduces a second layer of uncertainty, known as model (epistemic) uncertainty. In this paper, we propose an LSTM Autoencoder that quantifies both data and model uncertainty, enabling a deeper understanding of noise recognition. Our experimental results across different real-world datasets show that consideration of uncertainty effectively increases the robustness to noise and point outliers, making predictions more reliable for longer periodic sequential data.