HYDRA: Competing convolutional kernels for fast and accurate time series classification

Abstract

We demonstrate a simple connection between dictionary methods for time series classification, which involve extracting and counting symbolic patterns in time series, and methods based on transforming input time series using convolutional kernels, namely Rocket and its variants. We show that by adjusting a single hyperparameter it is possible to move by degrees between models resembling dictionary methods and models resembling Rocket. We present Hydra, a simple, fast, and accurate dictionary method for time series classification using competing convolutional kernels, combining key aspects of both Rocket and conventional dictionary methods. Hydra is faster and more accurate than the most accurate existing dictionary methods, and can be combined with Rocket and its variants to further improve the accuracy of these methods.

WEASEL 2.0 -- A Random Dilated Dictionary Transform for Fast, Accurate and Memory Constrained Time Series Classification

“…Yet, convolution filters are found from the entire space of possible real-values. The most well known approach is ROCKET [5], with its successors MiniROCKET [6], MultiROCKET [32], and Hydra [7]. ROCKET generates tens of thousands of randomly parameterized convolutional kernels, and applies two pooling operations to the output: the maximum and the proportion of positive values.…”

“…Dictionary (D): BOSS, cBOSS, WEASEL, TDE; Hybrid (H): HiveCote 2.0, HiveCote 1.0, TS_CHIEF, Deep-Learning (DL): InceptionTime; Shapelets (S): R-DST, MrSQM_SFA_k5; Kernel (K): Arsenal, MiniRocket, MultiRocket, Rocket, Hydra. We used implementations available in sktime [20], or published by the authors [7,12,17]. All reported numbers are accuracy on the test split.…”

“…Both in combination generate a low bias and low variance transformation usable by a linear classifier, for high accuracy and fast prediction. Another recent discovery is the use of a dilation operation applied to a filter, such as the convolutional filter (ROCKET [5][6][7]32]) or a sliding window operation (R-DST [12]). Dilation adds a gap between values, effectively increasing the size of the receptive field of the filter, and also operating similar to a down-sampling operation.…”

“…Figure 1 shows the advances made by WEASEL 2.0 in comparison to SotA dictionary transforms (Hydra [7] is a hybrid between kernelbased and dictionary, and MrSQM [17] is a hybrid between Shapelet and Dictionary). On 114 UCR datasets it is the best of its class, and significantly more accurate than its predecessor WEASEL, and furthermore it has a predictable memory footprint (see Figure 13).…”