Table 1 Comparative analysis of state-of-the-art time-series pattern recognition methods.
From: PatternFusion: a hybrid model for pattern recognition in time-series data using ensemble learning
Study | Approach | Strengths | Limitations | Performance metrics |
|---|---|---|---|---|
Hochreiter and Schmidhuber20 (LSTM) | Recurrent neural network with gated memory cells | Effective modeling of long-term dependencies; Robust to noise | Limited interpretability; High computational cost; No explicit multi-scale capability | Accuracy: 85–92% on sequence classification |
Vaswani et al.45(Transformer) | Self-attention mechanism without recurrence | Parallel computation; Direct modeling of arbitrary time relationships; Scalable | Quadratic complexity with sequence length; Limited interpretability; Requires large training datasets | Outperforms RNNs in language tasks; F1: 89–95% |
Dempster et al.11(ROCKET) | Random convolutional kernels with linear classifier | Exceptional computational efficiency; State-of-the-art accuracy on many benchmarks | Limited interpretability; No confidence measures; Static model integration | Classification accuracy: 92–96% on UCR archive |
Qin et al.40 (DA-RNN) | Dual-stage attention with RNN | Input feature selection; Temporal relevance weighting; Enhanced interpretability | Limited to forecasting tasks; No integration with statistical models; Single temporal scale | MSE 21–42% lower than baselines |
Li et al.33 (LSTM-SVR) | Hybrid statistical and deep learning | Combines LSTM memory with SVR generalization; Partial interpretability | Static integration of models; Limited to specific domains; No confidence measures | RMSE improved by 15–32% over single models |
Zhou et al.54 (Informer) | Efficient transformer with ProbSparse attention | Handles long sequences efficiently; Probabilistic attention mechanism | Focus on forecasting rather than pattern recognition; Limited interpretability; No statistical model integration | MSE reduced by 38–51% vs. traditional transformers |
Wu et al.49 (Autoformer) | Decomposition transformer with auto-correlation | Combines statistical decomposition with deep learning; Multi-scale architecture | Static integration strategy; Limited confidence quantification; High complexity | MSE improved by 9–23% over Informer on long sequences |
