MiniRocket: A Very Fast (Almost) Deterministic Transform for Time Series Classification

Daniel F. Schmidt,Geoffrey I. Webb,Angus Dempster

DOI: https://doi.org/10.1145/3447548.3467231

2020-12-16

Abstract:Rocket achieves state-of-the-art accuracy for time series classification with a fraction of the computational expense of most existing methods by transforming input time series using random convolutional kernels, and using the transformed features to train a linear classifier. We reformulate Rocket into a new method, MiniRocket. MiniRocket is up to 75 times faster than Rocket on larger datasets, and almost deterministic (and optionally, fully deterministic), while maintaining essentially the same accuracy. Using this method, it is possible to train and test a classifier on all of 109 datasets from the UCR archive to state-of-the-art accuracy in under 10 minutes. MiniRocket is significantly faster than any other method of comparable accuracy (including Rocket), and significantly more accurate than any other method of remotely similar computational expense.

Computer Science,Mathematics

What problem does this paper attempt to address?

The paper aims to address the issues of efficiency and accuracy in time series classification. Specifically, the researchers propose a new method called MiniRocket, which is an extremely fast and almost deterministic transformation method for time series classification tasks. Compared to the previous Rocket method, MiniRocket improves processing speed by up to 75 times on large datasets and is almost deterministic (or optionally fully deterministic), while maintaining nearly the same accuracy.

The Rocket method transforms the input time series using random convolutional kernels and trains a linear classifier with these transformed features. However, this method still has some randomness and computational complexity issues. Therefore, MiniRocket improves efficiency and determinism through the following enhancements:
1. **Changes in convolutional kernels**: MiniRocket uses a set of fixed small-scale convolutional kernels instead of randomly generated ones.
2. **Weight constraints**: The weights of MiniRocket's convolutional kernels are restricted to two values (-1 and 2), and their sum is zero.
3. **Bias selection**: Bias values are selected from the convolution output rather than being randomly generated.
4. **Dilation operation**: MiniRocket employs a fixed dilation rate, adjusted according to the input length.

These improvements make MiniRocket not only faster than Rocket but also competitive in terms of accuracy. It can complete the training and testing of all 109 datasets in the UCR archive within minutes, whereas Rocket requires several hours. Additionally, MiniRocket demonstrates significant advantages in terms of time and resource consumption, especially when handling large-scale datasets.