The Latent Behavior Space - A Vector Space for Time-Series Data

The timing and temporal order are two characteristic properties that are frequently omitted in machine learning approaches, but carry crucial information. Their consideration is currently limited to algorithms that are specialized to sequential data, but it takes a projection into a vector space to employ the wealth of ML algorithms that are known and understood. Projections inevitably cause a loss of detail. A naı̈ve application of bag-of-words, as a prominent example, utilizes neither order nor timing of events. In this paper we introduce a projection strategy that retains order and timings. It identifies the latent space that the generating processes underlying the time series are spanning.