Modeling time series by aggregating multiple fuzzy cognitive maps.

Area of Science:

• Time Series Analysis
• Computational Intelligence
• Data Science

Background:

• Real-world time series exhibit complex variations due to multiple influences, challenging single-model simulation accuracy.
• Existing time series prediction methods primarily focus on numerical and interval predictions, often neglecting model interpretability.
• A gap exists in developing time series models that are both accurate and easily understood by humans.

Purpose of the Study:

• To propose a novel prediction modeling methodology for time series that enhances interpretability.
• To address the limitations of existing methods in accurately reflecting time series variation characteristics.
• To develop a model that excels in numerical and interval predictions while being human-comprehensible.

Main Methods:

• A new prediction modeling methodology based on fuzzy cognitive maps (FCMs) is proposed.
• The bootstrap method is employed to select multiple sub-sequences capturing diverse variation modalities.
• FCMs are constructed for each sub-sequence and merged using granular computing for a comprehensive model.

Main Results:

• The established FCM-based model demonstrates strong performance in both numerical and interval predictions.
• The proposed approach significantly enhances the interpretability of time series prediction models.
• Experimental validation on synthetic and real-life datasets confirms the approach's usefulness and efficiency.

Conclusions:

• The proposed fuzzy cognitive map methodology offers a powerful and interpretable approach to time series prediction.
• This method effectively handles complex time series variations, providing accurate predictions and valuable insights.
• The integration of FCMs and granular computing represents a significant advancement in interpretable time series modeling.