MINIROCKET: A Very Fast (Almost) Deterministic Transform for Time Series Classification (2012.08791v2)

Abstract: Until recently, the most accurate methods for time series classification were limited by high computational complexity. ROCKET achieves state-of-the-art accuracy 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, making it up to 75 times faster on larger datasets, and making it almost deterministic (and optionally, with additional computational expense, 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 less than 10 minutes. MINIROCKET is significantly faster than any other method of comparable accuracy (including ROCKET), and significantly more accurate than any other method of even roughly-similar computational expense. As such, we suggest that MINIROCKET should now be considered and used as the default variant of ROCKET.

• The paper introduces a deterministic reformulation of the Rocket algorithm using a fixed set of convolution kernels to boost computational efficiency.
• The paper demonstrates that MiniRocket achieves up to a 75-fold speed improvement on large time series datasets while maintaining competitive accuracy.
• The paper details a method leveraging binary-like weights and PPV feature extraction to streamline convolution processes for real-time analysis.

An Examination of MiniRocket: A High-Speed Deterministic Approach for Time Series Classification

The paper under review details the reformulation of the Rocket algorithm into a new method named MiniRocket, which offers advancements in computational efficiency while maintaining comparable accuracy for time series classification tasks. This reformulation addresses the persistent challenge of high computational complexity often inherent in state-of-the-art methodologies for time series analysis.

Overview and Methodology

MiniRocket retains the core structure of Rocket by transforming input time series using convolutional kernels but achieves significant computational savings through a series of optimizations that make the method effectively deterministic. The principal innovation involves using a small, fixed set of kernels with weights restricted to two values, which allows the algorithm to sidestep the generation of random kernels characteristic of Rocket.

In particular, MiniRocket employs a precise, deterministic selection of kernel properties such as length, dilation, and padding. By focusing on kernels of length 9 and utilizing binary-like weights of {-1, 2}, the algorithm reduces the randomness involved in Rocket. Bias values are drawn directly from the convolution output, aligning with the input scale and negating the need for input normalization. This methodological reformulation of sampling bias boasts minimal randomness and, with an optional mode, achieves full determinism.

A noteworthy optimization strategy of MiniRocket is its ability to compute PPV (proportion of positive values) simultaneously for each kernel and its inverse, leveraging the inherent property of PPV’s bound between zero and unity. Transform time is dramatically reduced through the reuse of convolution outputs for feature extraction across multiple bias values. Furthermore, multiplication operations are avoided by precomputing and reusing values during the convolution process.

Performance and Accuracy

When evaluated on the extensive UCR time series classification archive, MiniRocket exhibits a remarkable computational advantage while maintaining competitiveness in accuracy with Rocket and other leading-edge time series classifiers like TS-CHIEF and HIVE-COTE/TDE. On average, MiniRocket is found to be marginally more accurate than Rocket, and significantly less computationally demanding, achieving up to a 75-fold faster performance on large datasets. This efficiency is underscored by the ability to operate within tight computational constraints—a stark contrast to the extensive time required by comparably accurate methods.

MiniRocket's deterministic nature does not compromise its accuracy; even in its fully deterministic form, it performs comparably to its default configuration using stochastic sampling of bias. Importantly, MiniRocket’s design enables it to exceed Rocket’s speed across datasets of varying sizes, transforming hundreds of thousands of time series efficiently and effectively.

Implications and Future Directions

The methodological innovations introduced by MiniRocket have profound implications for the fields of machine learning and data mining, especially in areas where time series data is prevalent. The significant reduction in computational costs opens avenues for real-time applications and deployments on large datasets without the prohibitive resource requirements traditionally associated with state-of-the-art time series classification methods.

The paper also hints at potential extensions of MiniRocket into multivariate time series analysis, nonlinear classification integration, and broader data transformation tasks. Such trajectories hold promise for enhancing the adaptability and applicability of time series classification techniques to more complex and diverse datasets, potentially reshaping the paradigms of how time series data is processed and analyzed in practical, real-world scenarios.

In conclusion, MiniRocket exemplifies a strategic step forward in balancing computational efficiency with robust classification accuracy, setting a precedent for future explorations in deterministic approaches for data-driven time series analysis.