Time series classification with random convolution kernels: pooling operators and input representations matter

Abstract Time series classification (TSC) is a fundamental problem in supervised learning, with applications in domains ranging from healthcare and finance to industrial monitoring. Recent advances based on random convolution kernels, such as ROCKET, MiniRocket, MultiRocket, and HYDRA, have demonstrated that large sets of randomly generated convolutional features combined with simple linear classifiers can achieve state-of-the-art accuracy at very low computational cost. These methods typically rely on fixed input representations and a small number of pooling operators, most commonly the proportion of positive values (PPV), to extract features from activation maps. However, the optimal choice of pooling operator and representation can vary substantially across datasets. In this work, we first provide an empirical analysis on the UCR archive showing that using a fixed pooling operator does not generally yield optimal performance, and that the best choice depends on the data. Motivated by this observation, we propose SelF-Rocket (Selected Features Rocket), a novel extension of MiniRocket that dynamically attempts to select the most appropriate combination of input representation and pooling operator during training. SelF-Rocket incorporates a feature selection module that adapts to each dataset, allowing it to select the most appropriate pooling operator and representation. Extensive experiments on the UCR benchmark demonstrate that SelF-Rocket achieves competitive or superior accuracy compared with existing random convolution kernel methods.