From idealized to realistic settings : deep learning-based imputation for multivariate time series

With the increasing collection of large-scale data and the rapid development of machine learning (ML) techniques, industries across various domains are striving to automate processes and make more informed, data-driven decisions. However, many existing systems were not originally designed with ML applications in mind, resulting in datasets that frequently contain missing values due to inconsistent data collection practices, legacy infrastructure, or context-dependent omissions. These challenges are particularly pronounced in multivariate time series data, where missingness can disrupt both temporal and inter-feature dependencies. While imputation techniques are commonly used to address this, most evaluations rely on simplified assumptions that do not reflect the structured or biased missingness often observed in real-world data. This thesis investigates deep learning–based methods for imputing missing values in multivariate time series, with a focus on performance under realistic missingness conditions. It explores two complementary directions: controlled benchmarking and real-world evaluation in a process mining context using regulatory workflow data. The first direction involves a comparative study of state-of-the-art recurrent and attention-based models across three datasets, multiple missingness mechanisms, and a range of missing rates. In total, 288 experiments assess imputation quality, with 194 additional configurations evaluating downstream binary classification. The second direction applies state-of-the-art deep imputation methods to a case study using data from the British Columbia Energy Regulator (BCER), where implausibly short task duration records pose challenges for predictive modeling. Together, the results offer a realistic assessment of deep learning–based imputation methods. While the observed benefits were less pronounced than those reported under idealized evaluation settings, deep imputation models nonetheless demonstrated their value in handling missing or corrupted time series data, consistently improving data quality and downstream prediction.