Source-linked AI summary
ROCKET: Exceptionally fast and accurate time series classification using random convolutional kernels
Angus Dempster, François Petitjean, Geoffrey I. Webb
TL;DR
High-accuracy time series classifiers can be computationally expensive and may focus on limited feature types. Rocket addresses this by transforming series with many random convolutional kernels and training a linear classifier, achieving state-of-the-art accuracy with a fraction of existing methods’ computational expense and scaling to millions of series.
Problem
High-performing time series classifiers often have high computational complexity, require substantial training time, and may not scale to large datasets or capture diverse feature types with one method.
Method
Rocket transforms time series using many random convolutional kernels and trains a linear classifier on the transformed features.
Results
Rocket achieves state-of-the-art accuracy with a fraction of existing methods’ computational expense and scales to millions of time series.
Takeaways & Limitations
Random convolutional kernels collectively capture discriminative time-series patterns while requiring low computational resources.
Takeaways & Limitations
Rocket was developed on 40 randomly selected datasets from the 85-dataset collection, with the remaining datasets reserved as holdouts.
Abstract
from arXiv · showhide
Most methods for time series classification that attain state-of-the-art accuracy have high computational complexity, requiring significant training time even for smaller datasets, and are intractable for larger datasets. Additionally, many existing methods focus on a single type of feature such as shape or frequency. Building on the recent success of convolutional neural networks for time series classification, we show that simple linear classifiers using random convolutional kernels achieve state-of-the-art accuracy with a fraction of the computational expense of existing methods.
1 Introduction
Rocket addresses the computational cost and limited feature focus of high-performing time series classifiers by combining random convolutional kernels with a linear classifier. It achieves state-of-the-art accuracy while substantially reducing computation and scaling to large datasets.
- Motivation: Existing state-of-the-art time series classifiers often require substantial training time and do not scale to large datasets.This computational burden has motivated more scalable methods, including Proximity Forest, TS-CHIEF, and InceptionTime.
- Method: Rocket transforms time series with random convolutional kernels and trains a linear classifier on the resulting features.The kernels vary in length, weights, bias, dilation, and padding.
- Method: Random kernels collectively capture classification-relevant features despite individual kernels only approximately capturing relevant patterns.This provides a single mechanism for features that previously required specialized techniques.
- Results: Rocket achieves state-of-the-art accuracy on the 85 UCR ‘bake off’ datasets with substantially less training time than existing methods.Figure 1 compares mean ranks, while the reported timings show Rocket requiring only a fraction of competing methods’ training time.
- Results: 1 hour 50 minutes is Rocket’s total compute time across all 85 datasets, versus more than 6 days for InceptionTime and more than 11 days for Proximity Forest and TS-CHIEF.Rocket timings used one CPU core per dataset per run and averaged 10 runs.
- Scalability: 1 hour 15 minutes is Rocket’s training time for 1 million time series, compared with more than 16 hours for Proximity Forest at similar accuracy.A restricted Rocket variant takes less than 1 minute with slightly lower accuracy.
2 Related Work
Time series classification methods span feature-specific representations, scalable ensembles, and convolutional approaches, while computational complexity remains a central constraint. Rocket builds on convolutional kernels by combining varied random architectures and extracted features to target accuracy with lower computational expense.
- Existing classifiers commonly specialize in frequency, variance, shapelets, elastic distances, or combinations of these representations.
- BOSS and Shapelet Transform have training complexities of O(n^2 · l^2) and O(n^2 · l^4), respectively.
- HIVE-COTE combines dictionary, shapelet, elastic-distance, and frequency-based classifiers but inherits high computational complexity.
- Proximity Forest and TS-CHIEF improve scalability through tree-based splitting, with quasilinear dependence on examples but quadratic dependence on series length.
- Convolutional kernels detect patterns directly and can capture shapes, multiple scales, frequencies, positions, and warping through kernel combinations, dilation, pooling, and randomization.
- Rocket distinguishes itself by leveraging varied random kernel lengths, dilations, padding, weights, biases, and output features for accurate, computationally efficient classification.
3 Method
Rocket transforms time series with many random convolutional kernels, aggregates each resulting feature map, and trains a linear classifier on the transformed features. Its design combines broad random kernel variation with efficient computation and feature extraction for scalable classification.
- 3 Method: Rocket uses random kernels with varied length, weights, bias, dilation, and padding to transform time series into features.Kernel parameters are randomly generated, with configurations selected through development-set experimentation.
- 3 Method: Rocket uses a linear classifier on transformed features, with ridge regression used for UCR experiments and logistic regression with Adam for scalability studies.Ridge regression has effective complexity O(n^2·f) or O(n·f^2), depending on the decomposition used.
- 3 Method: A single layer of many unlearned kernels keeps convolutional computation inexpensive while providing substantial kernel variety.The method uses a very large number of kernels and relies on their combination to capture classification-relevant features.
- 3 Method: The number of kernels k is Rocket’s effective hyperparameter, trading classification accuracy against transform time that grows linearly with k.The default configuration uses 10,000 kernels while remaining extremely fast.
- 3 Method: Each kernel produces a feature map, from which Rocket extracts the maximum value and the proportion of positive values, or ppv.The ppv captures how much of the input matches a pattern and performed better than alternatives such as the mean.
- 3 Method: The transform applies dilated convolutions with randomly selected padding, while dilation expands effective kernel size and can reduce computation without padding.Dilation enables similar kernels to match patterns at different frequencies and scales; padding is selected with equal probability.
4 Experiments
Rocket is evaluated for accuracy and scalability across UCR datasets, variable-length series, training-set sizes, and series lengths. It remains competitive with leading classifiers while substantially reducing computational time.
- UCR archive: Rocket obtains the best mean rank over the 85 UCR ‘bake off’ datasets.The comparison includes BOSS, Shapelet Transform, Proximity Forest, ResNet, HIVE-COTE, InceptionTime, and TS-CHIEF.
- UCR archive: Rocket ranks slightly ahead of HIVE-COTE, TS-CHIEF, and InceptionTime, but their accuracy differences are not statistically significant.TS-CHIEF ranks ahead of Rocket on the 45 holdout datasets, also without a significant difference.
- Training set size: With 1,000 or 10,000 kernels, Rocket achieves similar accuracy to Proximity Forest and TS-CHIEF, while even 10,000 kernels are an order of magnitude faster than Proximity Forest.With 100 kernels, Rocket learns from more than 1 million time series in less than a minute, at lower accuracy.
- Time series length: Rocket takes approximately as long on length-2,048 series as TS-CHIEF takes on length-32 series and is approximately three orders of magnitude faster at the longest length.Rocket is also considerably faster than InceptionTime, although their fundamental scalability is likely similar because both use convolutional architectures.
4.3 Sensitivity Analysis
Sensitivity experiments show that Rocket is robust to many parameter choices, while dilation and ppv are especially important for classification accuracy.
- Overall sensitivity: Alternative configurations usually produce accuracy similar to the default, indicating that Rocket is relatively robust to many parameter choices.The analysis varies kernel count, length, weights, bias, dilation, padding, centering, and output features.
- Number of kernels: k = 10,000 is ahead of k = 5,000 in win/draw/loss at 29/3/8, while differences among k = 10,000, 50,000, and 100,000 are not statistically significant.Accuracy variability also decreases as kernel count increases: median standard deviation is 0.0038 at k = 10,000 and 0.0021 at k = 100,000.
- Kernel length: Kernel lengths of 7, 9, or 11, and random selection from {5, 7, 9} or {9, 11, 13}, match the default accuracy without significant differences.Shorter kernels are undesirable because they are more strongly correlated when many kernels are used.
- Dilation: Dilation produces higher accuracy than no dilation, and exponential dilation is significantly better than uniform dilation on some datasets.Dilation is identified as a key aspect of Rocket’s performance.
- Padding: Padding is superior to no padding, but the differences among padding choices are not statistically significant.Random, uniform, and centered padding produce very similar results.
- Output features: ppv produces substantially higher accuracy than max for most development datasets and has the single biggest effect among the evaluated parameters.Combining ppv and max improves accuracy further, although the additional difference is small.
5 Conclusion
Rocket combines many varied random convolutional kernels with low-cost feature extraction and linear classification to achieve accurate, scalable time series classification. Its ppv feature summarizes pattern prevalence more effectively than maximum pooling in these experiments.
- Contribution: Large numbers of random kernels jointly capture discriminative time-series patterns despite individual kernels only approximately matching relevant features.Rocket uses random kernels rather than learned kernel weights.
- Contribution: Rocket achieves state-of-the-art accuracy with a fraction of existing methods’ computational expense and scales to millions of time series.The method’s low computational requirements make both learning and classification extremely fast.
- Feature representation: The proportion of positive values, or ppv, summarizes feature-map outputs by measuring pattern prevalence in a time series.The authors report that ppv is substantially more effective than a simple maximum and, to their knowledge, had not previously been used this way.
- Future work: Future work includes feature selection, multivariate time series, applications beyond time series, and combining aspects of Rocket with learned kernels.
A Relative Accuracy
Figure 13 presents the relative accuracy of Rocket compared with state-of-the-art classifiers on the “bake off” datasets.
- Figure 13 compares Rocket’s relative accuracy with state-of-the-art classifiers across the “bake off” datasets.
B ‘Development’ and ‘Holdout’ Datasets
Figures 14 and 15 report Rocket’s mean rank against state-of-the-art classifiers on holdout and development datasets.
- Figure 14 compares Rocket’s mean rank with state-of-the-art classifiers on the holdout datasets.
- Figure 15 compares Rocket’s mean rank with state-of-the-art classifiers on the development datasets.
C Additional Plots
Figures 16–23 examine relative accuracy across alternative Rocket design choices on the development datasets.
- Figure 16 compares relative accuracy for k = 10,000 versus k = 5,000.
- Figures 17–20 compare relative accuracy across kernel lengths, weight distributions, centering strategies, and bias distributions.
- Figures 21–23 compare relative accuracy across dilation, padding, and feature-aggregation choices.
D Results for ‘Bake Off’ Datasets
This section presents classification accuracy results for the ‘Bake Off’ datasets, with development datasets marked by an asterisk.
- Development datasets are marked with an asterisk.
- Table 1 reports classification accuracy for the ‘Bake Off’ datasets.
E Results for Additional 2018 Datasets
Table 2 reports classification accuracy results for additional 2018 datasets.
- Table 2 presents classification accuracy for the additional 2018 datasets.