Original title: Time series classification with random convolution kernels: pooling operators and input representations matter
Mouhamadou Mansour Lo, Gildąs Morvan, Mathieu Rossi, Fabrice Morganti +1 · September 2026
Classifying time series data, like tracking health stats or factory machines over time, is an important challenge in machine learning. Popular, fast tools like MiniRocket scan this data using random pattern matchers called convolution kernels. They then summarize these matches into simple numbers using fixed rules called pooling operators. Most tools use just one fixed rule and one fixed way to look at the data. But the best rule changes depending on the dataset. To fix this, researchers created SelF-Rocket. Instead of sticking to one fixed choice, SelF-Rocket learns during training to pick the best summary rule and data view for each specific dataset. It tests different options and selects the most useful features. The researchers found that SelF-Rocket achieves competitive or superior accuracy compared to existing random convolution kernel methods across standard benchmark tests.
Imagine you are trying to identify a song by listening to a few seconds. You have a set of rules to describe what you hear: the tempo, the instruments, the melody. But you always use the same rule, like counting the number of drum beats, even though some songs are better identified by their melody. Here, the time series is the song, the random patterns are different ways to listen, and the pooling operator is the rule you use to describe what you heard. SelF-Rocket picks the best rule for each song automatically.
Time series data appears in heart monitoring, stock prices, and factory sensors. Existing fast methods use a fixed way to summarize features, but that choice can limit accuracy. SelF-Rocket automatically picks the best summary for each dataset, improving results without slowing down.
The study is limited to the UCR archive and comparisons only with other random convolution kernel methods. It does not analyze the computational overhead of the feature selection module, nor does it test on other types of time series data or against non-convolutional methods.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Time series classification with random convolution kernels: pooling operators and input representations matter — Mouhamadou Mansour Lo, Gildąs Morvan, Mathieu Rossi et al. — Data Mining and Knowledge Discovery — doi:10.1007/s10618-026-01269-w
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.