Decentralized Time Series Classification with ROCKET Features

TL;DR

The paper tackles privacy-preserving time series classification (TSC) in distributed settings where server-based federated learning can be a single point of failure. It introduces DROCKS, a fully decentralized FL framework that extends FROCKS to multiclass tasks by employing a ring topology and sequential kernel selection for ROCKET features, with kernel transfers governed by $p = \left \lfloor \frac{K}{N} \right \rfloor$. Each federation node refines a shared linear model using a mix of received kernels and newly generated ROCKET kernels, and forwards the model along the ring while communicating only kernel seeds to minimize bandwidth; convergence is defined by stabilization of the top-$p$ kernel set across rounds. Evaluated on 128 UCR binary and multiclass datasets, DROCKS generally outperforms server-based FL baselines in multiclass scenarios, achieves lower communication and computation overhead, and demonstrates resilience to node failures and topology changes. The work also positions FROCKS as a binary-task precursor and discusses fault-tolerance strategies, scalable topology alternatives, and future improvements for large kernel counts and non-i.i.d. data settings, with practical implications for privacy-preserving, resource-efficient TSC across distributed devices.

Abstract

Time series classification (TSC) is a critical task with applications in various domains, including healthcare, finance, and industrial monitoring. Due to privacy concerns and data regulations, Federated Learning has emerged as a promising approach for learning from distributed time series data without centralizing raw information. However, most FL solutions rely on a client-server architecture, which introduces robustness and confidentiality risks related to the distinguished role of the server, which is a single point of failure and can observe knowledge extracted from clients. To address these challenges, we propose DROCKS, a fully decentralized FL framework for TSC that leverages ROCKET (RandOm Convolutional KErnel Transform) features. In DROCKS, the global model is trained by sequentially traversing a structured path across federation nodes, where each node refines the model and selects the most effective local kernels before passing them to the successor. Extensive experiments on the UCR archive demonstrate that DROCKS outperforms state-of-the-art client-server FL approaches while being more resilient to node failures and malicious attacks. Our code is available at https://anonymous.4open.science/r/DROCKS-7FF3/README.md.

Figures (10)

• Figure 1: Each node in the sequence receives the trained model and the $p$ best-performing kernels from the preceding node. The node then fine-tunes the received model using its private data, transformed with a new set of ROCKET kernels that combine new kernels with the received ones.
• Figure 2: Mean ranks for the competitors and DROCKS on binary classification tasks, each with the best hyperparameter.
• Figure 3: Mean ranks for the competitors and DROCKS on multiclass classification tasks, each with the best hyperparameters.
• Figure 4: Mean ranks of DROCKS with different kernel counts across all UCR datasets.
• Figure 5: Number of rounds until convergence shown over all datasets. The x-axis indicates the number of kernels used to initialize.