Double Decomposition and Fuzzy Cognitive Graph-Based Prediction of Non-Stationary Time Series

Area of Science:

• Time Series Analysis
• Machine Learning
• Signal Processing

Background:

• Deep learning models like Recurrent Neural Network (RNN) excel at time series forecasting but lack interpretability, hindering decision-maker trust.
• Interpretable and accurate prediction models are crucial for reliable decision-making in complex systems.

Purpose of the Study:

• To develop an interpretable and accurate time series prediction model.
• To enhance decision-maker trust in forecasting models through explainability.

Main Methods:

• A novel WE-HFCM model is proposed, integrating Wavelet Decomposition (WD) and Empirical Mode Decomposition (EMD) for signal decomposition.
• High-order Fuzzy Cognitive Maps (HFCM) are constructed for interpretability and reasoning.
• Ridge regression is employed to learn the HFCM weight vector for predictive modeling.

Main Results:

• The WE-HFCM model demonstrated superior prediction accuracy on both stationary and non-stationary datasets.
• For stationary series, WE-HFCM accuracy was 45% higher than ARIMA, 35% higher than SARIMA, and 16% higher than LSTM.
• For non-stationary series, WE-HFCM accuracy surpassed ARIMA and SARIMA by 69%.

Conclusions:

• The proposed WE-HFCM model effectively combines decomposition techniques with interpretable fuzzy cognitive maps for accurate time series forecasting.
• The model offers a significant improvement in prediction accuracy and interpretability over existing methods like ARIMA and LSTM.