Advanced Time Series Embeddings Using Prior Knowledge of Change Point Information

This paper presents an advanced framework for enhancing time series segment representation through the in-corporation of change point information. The proposed method utilizes multiple state-of-the-art embedding deep neural networks (DNNs), including a custom autoencoder, Deep Temporal Convolution Network (DTCN), and Time-series Dense Encoder (TiDE). Central to our approach is the introduction of the intra-and inter-segment loss (I2SL) component, a novel loss function designed to leverage change point information. The I2SL component is integrated into the training process of the embedding DNNs, where it serves two key roles: (1) intra-segment coherence, ensuring that embeddings of time windows within the same segment remain similar, and (2) inter-segment differentiation, penalizing large similarities of embeddings from adjacent segments that are separated by a detected change point. The I2SL, as dual-objective function is critical in promoting embeddings that not only capture the intrinsic properties of the time series but also respect the temporal boundaries indicated by change points. The efficacy of our approach is evaluated on a diverse set of multiclass datasets from the Time Series Segmentation Benchmark (TSSB), using various clustering algorithms including K-means, DBSCAN, and hierarchical. Our experimental results demonstrate that the I2SL component significantly improves the clarity and separation of clusters, making it a valuable method for time series analysis and unsupervised learning applications. The code implementation for this study is available on GitHub at https://github.com/marcell-nemeth/ts_embeddings_chp.