Time Series Representation Learning

Time series representation learning studies how to encode sequences so that useful patterns—such as trends, periodicity, local events, and relationships among channels—are easier to recognize or predict. Approaches range from symbolic and signature-based features to learned patch, frequency, and graph representations, often trained with contrastive, masked-prediction, or reconstruction objectives. Pretrained time series models and multimodal methods aim to transfer representations across datasets and tasks, including medical signals, sensor data, and irregularly sampled sequences. A central challenge is that general-purpose representations do not always transfer to specialized or shifted settings, and strong prediction alone does not ensure that a representation preserves interpretable or causally relevant structure.