RACER: Fast and Accurate Time Series Clustering With Random Convolutional Kernels and Ensemble Methods

The widespread adoption of sensor devices and IoT technologies has generated vast amounts of time series data, making time series data mining a key research focus for informed decision-making in IoT. This paper focuses on time series clustering, a key task that has driven significant research and algorithm development. Despite progress, current methods still offer room for improvement in accuracy and efficiency. To this end, this paper presents RACER, a time series clustering method that achieves high accuracy and efficiency by bridging the gap between random convolutional kernels and ensemble clustering. Specifically, RACER utilizes random convolutional kernels to extract features from raw data and obtains six base clustering results by combining various dimensionality reduction techniques with clustering methods tailored for the extracted features. The final clustering results are derived through ensemble clustering using non-negative matrix factorization. We evaluate RACER on the UCR Archive using the Adjusted Rand Index. Experimental results reveal that: (1) RACER surpasses recently proposed state-of-the-art algorithms, including R-Clustering and RandomNet, as well as classical methods such ask-Shape andk-means with DTW; (2) RACER completes clustering on 128 UCR datasets (excluding 11 datasets with unequal lengths) in just 18 minutes, significantly outperforming the vast majority of existing approaches in terms of efficiency. The experimental results and the code for RACER are available at the GitHub link: https://github.com/pcjdhhhh/RACER.