<ul><li>Existing time series tokenization methods encode a constant number of samples into individual tokens, resulting in computational overhead.</li><li>A new pattern-centric tokenization scheme for time series analysis is proposed, based on a discrete vocabulary of frequent motifs.</li><li>The method merges samples with underlying patterns into tokens, compressing time series adaptively.</li><li>The motif-based tokenization improves forecasting performance by 36% and boosts efficiency by 1990% on average.</li></ul>

Byte Pair Encoding for Efficient Time Series Forecasting

Discover more