Data collected from real-world applications can usually be modeled as time-dependent observations that form multivariate time series, e.g., energy production in different locations. Time series analytics plays an important role in ensuring the effective functionality of applications in different scenarios. For example, time series forecasting can help improve schedules in transport systems, and anomaly detection can help improve safety and reliability in computing systems. Recent works with deep learning technologies have demonstrated good performance in multivariate time series analytics, but these approaches are usually based on data-driven models. They struggle to utilize real-world knowledge, which can enhance reliability to improve accuracy for data-driven models.

This thesis addresses this by three contributions:

1. MTSF-DG – we incorporate dynamic graph modeling and causality knowledge into multivariate time series forecasting, where correlations among variates can differ across time.
2. TAD-UP – we incorporate continuity and discreteness knowledge into anomaly detection, where different variates may change continuously or jump across time.
3. IGCL – we explore unsupervised time series anomaly prediction using Maximum Mean Discrepancy (MMD) knowledge, to predict anomalies for timely maintenance of systems.

Together, these contributions improve deep tme series analytics with real-world knowledge. 

All interested parties are welcome. After the defence the department will be hosting a small reception in cluster 3.