Many-to-Many Prediction for Effective Modeling of Frequent Label Transitions in Time Series

Alexander Katrompas, Vangelis Metsis

Abstract: Time-series classification is vital in health monitoring and human activity recognition, as well as in areas such as financial forecasting, process control, and a wide array of forecasting tasks. Traditional time-series models segment data into windows and assign one label per window, often missing label transitions within those windows. This paper presents a novel many-to-many time-series model and post-processing using hybrid recurrent neural networks with attention mechanisms, which more effectively captures label transitions over traditional many-to-one models. Further, unlike typical other many-to-many models, our approach doesn’t require a decoder. Instead, it employs an RNN, generating a label for every input time step. During inference, a weighted voting scheme consolidates overlapping predictions into one label per time step. Experiments show our model remains effective on time-series with sparse label shifts, but particularly excels in detecting frequent transitions. This model is ideal for tasks demanding accurate pinpointing of rapid label changes in time-series data, such as gesture recognition, making it ideal for fast-paced human activity recognition. 1

External IDs:dblp:conf/petra/KatrompasM24