Dimensionality reduction of time-series data, and systems and devices that use the resultant embeddings

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for dimensionality reduction of time-series using contrastive learning. A method can include receiving multidimensional input time series data that includes data from a session or sessions that span time, selecting positive samples and negative samples from the multidimensional input time series data for respective reference samples from the multidimensional input time series data, wherein the positive samples and negative samples are each selected with a respective predetermined distribution across the time of the session or sessions of the multidimensional input time series, and mapping the reference samples, the positive samples, and negative samples into a common embedding space.