Early time-series classification with SPRT and normalizing flow

This paper proposes a novel approach for early time-series classification, addressing the trade-off between prediction accuracy and earliness, which is critical in real-time applications. The Sequential Probability Ratio Test (SPRT) provides an optimal solution but relies on prior knowledge of the data’s probability distribution, which is an assumption often impractical in real-world scenarios. Existing studies commonly assume a normal distribution, which limits classification performance in complex data. To overcome this limitation, we integrate normalizing flow into the SPRT framework, enabling the estimation of conditional probability distributions through a series of invertible transformations. This approach allows for precise probability estimation, improving the accuracy of early classification. Experimental results on a preprocessed dataset demonstrate that the proposed model significantly enhances classification performance, offering a promising direction for advancing early time-series classification.