Principled change point detection via representation learning for data of different modalities

In sequential data analysis, a change point is a moment of abrupt regime switch in data streams. Detecting these changes quickly is crucial in various scenarios, from industrial sensor monitoring to challenging video surveillance. Classic approaches for change point detection (CPD) often struggle with high-dimensional semi-structured data due to their intricate nature, as they cannot process such data without a proper representation. Moreover, they can suffer from strong assumptions and lack of domain-specificity involvement even when applied to simpler numerical time series. In this thesis, we propose a set of solutions for CPD via representation learning. A principled loss function that allows representation learning for deep models is presented for the first time. Our loss balances change detection delay and time to a false alarm. It approximates classic rigorous solutions, while being differentiable unlike them. We also study real industrial problems from the oil&gas sector, demonstrating how to derive domain-specific embeddings for CPD tasks. Through comprehensive experiments on synthetic sequences, real-world sensor data, and video datasets, we demonstrate the importance of meaningful representations tailored to the CPD task. Notably, our results indicate that even data with a straightforward structure can benefit from utilizing proper embeddings designed to distinguish between similar and dissimilar observations. Our solutions outperform existing baselines, with an F1 score of $0.53$ for explosion detection in video, compared to baseline scores of $0.31$ and $0.35$. This thesis provides a robust methodology for CPD in various data modalities, offering significant improvements over traditional approaches and contributing valuable insights and tools for future research.