TLDR: HIT-ROCKET is a novel time series classification algorithm that uses Hadamard vectors as convolutional kernels to significantly improve computational efficiency and classification accuracy. It achieves state-of-the-art performance with at least a 5% F1-score improvement over ROCKET and 50% faster training than miniROCKET, making it ideal for deployment on ultra-low-power embedded devices. The method also demonstrates enhanced noise robustness and strong performance in one-class classification tasks.
Time series classification is a critical area with wide-ranging applications, from communications and finance to medicine. However, many existing state-of-the-art methods, while powerful, often come with significant computational demands, long training times, and complex parameter tuning. On the other end of the spectrum, lightweight solutions like ROCKET (Random Convolutional Kernel Transform) offer efficiency but still leave room for improvement in areas such as kernel selection and overall computational complexity.
Addressing these challenges, researchers from the Harbin Institute of Technology have introduced a novel feature extraction method called HIT-ROCKET. This innovative approach is based on the Hadamard convolutional transform, utilizing column or row vectors of the Hadamard matrix as convolution kernels. By extending convolution lengths and leveraging the inherent orthogonality of these Hadamard kernels, HIT-ROCKET not only maintains full compatibility with existing ROCKET-family methods but also significantly boosts computational efficiency, robustness, and adaptability.
The core idea behind HIT-ROCKET is to replace the ‘arbitrarily specified’ convolutions found in other ROCKET methods, such as the fixed kernels in miniROCKET, with Hadamard vectors. These vectors, composed solely of +1 or -1, offer strong mathematical support and enhance computational efficiency. A key advantage is the low correlation between these kernels, which improves classification accuracy and resistance to noise. In its main version, HIT-ROCKET primarily uses the Proportion of Positive Values (PPV) as a key feature, and its low feature correlation allows for excellent performance with various classifiers, including logistic regression, tree-based models, and Support Vector Machines (SVM).
Extensive experiments were conducted on the large-scale open-source University of California, Riverside (UCR) time series dataset, which covers diverse domains. The results are compelling: HIT-ROCKET achieves state-of-the-art performance across all tested datasets. Its F1-score, a measure of accuracy, improved by at least 5% compared to the original ROCKET method. Furthermore, when hyperparameters were configured identically, HIT-ROCKET’s training time was 50% lower than that of miniROCKET, which is known as the fastest variant in the ROCKET family. This remarkable efficiency makes HIT-ROCKET particularly well-suited for deployment on ultra-low-power embedded devices, such as microcontrollers and FPGAs.
The paper highlights several key contributions, including the mathematical derivation of performance improvements, comprehensive experimental validation on the UCR dataset, and the provision of additional feature extraction schemes like symplectic geometric reconstruction and non-uniform dilated convolution. The researchers have also made their code openly available on GitHub, ensuring that other researchers can easily apply the model and reproduce the experiments.
HIT-ROCKET’s advantages extend to noise robustness. Time series data often contains significant Gaussian noise. The method demonstrates more stable performance and smaller variance in F1 scores across different Signal-to-Noise Ratio (SNR) levels compared to miniROCKET, indicating its superior ability to handle noisy conditions. This is attributed to Hadamard convolution distributing signal energy across orthogonal channels, leading to smaller fluctuations in thresholded features.
Beyond standard classification, HIT-ROCKET also shows promise in one-class classification and anomaly detection tasks. When combined with robust classifiers like RLS-OCSVM, it achieved excellent recall metrics, further demonstrating its versatility and efficiency. The research also explored feature-enhanced forms of HIT-ROCKET, incorporating techniques like increasing dilation convolutions and differential sequence processing, which showed performance comparable to or exceeding MultiROCKET, another advanced ROCKET variant.
Also Read:
- Unlocking Robust Information Retrieval in Vector Symbolic Architectures
- OSBAD: A New Open-Source Benchmark for Battery Anomaly Detection
In conclusion, HIT-ROCKET represents a significant advancement in time series classification. By leveraging Hadamard vectors for convolution, it offers a powerful, efficient, and robust solution that is ideal for edge computing scenarios and embedded deployments. For more technical details, you can refer to the full research paper: HIT-ROCKET: Hadamard-vector Inner-product Transformer for ROCKET.
