Introduction
Multivariate Time Series Classification (MTSC) is a fundamental task in artificial intelligence, with applications spanning medical diagnostics (e.g., EEG analysis) and industrial monitoring. Despite its importance, real-world time series data presents significant modeling challenges due to multi-periodicity and scale-dependent correlations. Multi-periodicity refers to overlapping patterns repeating at different intervals, while scale-dependent correlations imply that the relationship between variables can shift—for instance, from positive to negative—depending on the temporal scale being analyzed.
Existing models, such as standard Convolutional Neural Networks (CNNs) or Transformers, often fail to capture these nuances. They typically focus either on local windows or global sequences, missing the underlying periodic structure that defines many complex signals.
Methodology: The MPTSNet Architecture
The proposed MPTSNet (Multiscale Periodic Time Series Network) addresses these limitations by explicitly modeling periodic structures. The architecture moves away from linear sequence processing, instead decomposing signals into multiple periodic scales.
1. Main Periods Identification
MPTSNet utilizes the Fast Fourier Transform (FFT) to identify the top $k$ most significant frequencies within the training data. These frequencies serve as a structural prior, guiding the network's feature extraction process across different scales.
2. The PeriodicBlock
The PeriodicBlock is the central innovation of MPTSNet. For each identified period, the 1D time series is reshaped into a 2D representation (Time $\times$ Period). This allows the model to apply two specialized modules:
- Local Extractor: An Inception-based module employing parallel 1D convolutions with varying kernel sizes (1 to 11). This captures "intra-period" local variations and shapes.
- Global Capturer: A multi-head attention mechanism that models "inter-period" dependencies, enabling the network to understand how patterns in one cycle relate to those in others.
3. Adaptive Aggregation
Recognizing that different periodic scales carry varying levels of importance, MPTSNet employs an adaptive fusion strategy. It uses the FFT amplitudes as weights, processed via a Softmax function, to aggregate features from different scales before they reach the final classification head.
Experimental Evaluation
MPTSNet was rigorously evaluated using the UEA multivariate time series classification archive. The experiments were divided into two settings:
- General Framework Comparison: MPTSNet was compared against 11 general time series analysis frameworks (e.g., TimesNet, PatchTST) on 10 diverse datasets. It achieved an average accuracy of 75.4%, ranking first overall.
- MTSC-Dedicated Comparison: Against 10 specialized MTSC models (e.g., Rocket, TapNet) across 25 datasets, MPTSNet maintained the highest average accuracy (74.0%) and the lowest average rank (3.12).
Discussion and Key Insights
A critical finding from the ablation studies was the importance of multiscale periodic decomposition. Removing this component resulted in a significant accuracy drop (from 75.4% to 70.9%), confirming that the periodic framework is essential for the model's success. Furthermore, the sequential application of local extraction followed by global capturing was found to be more effective than parallel processing, as it better manages redundancy within the periodic structure.
While MPTSNet demonstrates state-of-the-art performance, the selection of the hyperparameter $k$ (number of scales) remains a consideration. While $k=5$ proved robust for most benchmarks, future work may explore adaptive $k$-selection methods to automatically optimize the number of scales for individual datasets or samples.
Summary updated