Abstract
This paper presents a novel neural time series classifier that operates on symbolic data representation. The proposed approach extends an embedding-based model called SAFE (Simple And Fast segmented word Embedding) by introducing a dynamic variant in which the neural network architecture changes during training. The design of the method utilizes a learning scheme called GrowingNN, which adapts neural architectures through Monte Carlo Tree Search (MCTS). In the proposed hybrid model, SAFE converts time series into symbolic words and maps them to dense embeddings. Then, GrowingNN classifies the embedding matrices, starting from a small network and dynamically growing or shrinking it during training using MCTS-guided structural modifications. On 14 benchmark datasets, the combined approach achieves results comparable to ResNet and standalone SAFE. Empirical experiments show that the new method constructs very compact models tailored to the dataset, where the smallest network has only 1545 trainable parameters it on average three times smaller than ROCKET and 72 times smaller than that of ResNet.
Paper Type
Short Paper
DOI
10.62036/ISD.2026.49
Automated Neural Structure Adaptation for Time Series Classification
This paper presents a novel neural time series classifier that operates on symbolic data representation. The proposed approach extends an embedding-based model called SAFE (Simple And Fast segmented word Embedding) by introducing a dynamic variant in which the neural network architecture changes during training. The design of the method utilizes a learning scheme called GrowingNN, which adapts neural architectures through Monte Carlo Tree Search (MCTS). In the proposed hybrid model, SAFE converts time series into symbolic words and maps them to dense embeddings. Then, GrowingNN classifies the embedding matrices, starting from a small network and dynamically growing or shrinking it during training using MCTS-guided structural modifications. On 14 benchmark datasets, the combined approach achieves results comparable to ResNet and standalone SAFE. Empirical experiments show that the new method constructs very compact models tailored to the dataset, where the smallest network has only 1545 trainable parameters it on average three times smaller than ROCKET and 72 times smaller than that of ResNet.
Recommended Citation
Jastrzebska, A. & Świderski, S.(2026). Automated Neural Structure Adaptation for Time Series Classification. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.49