Modeling time series by aggregating multiple fuzzy cognitive maps

Background: The real time series is affected by various combinations of influences, consequently, it has a variety of variation modality. It is hard to reflect the variation characteristic of the time series accurately when simulating time series only by a single model. Most of the existing methods focused on numerical prediction of time series. Also, the forecast uncertainty of time series is resolved by the interval prediction. However, few researches focus on making the model interpretable and easily comprehended by humans. Methods: To overcome this limitation, a new prediction modelling methodology based on fuzzy cognitive maps is proposed. The bootstrap method is adopted to select multiple sub-sequences at first. As a result, the variation modality are contained in these sub-sequences. Then, the fuzzy cognitive maps are constructed in terms of these sub-sequences, respectively. Furthermore, these fuzzy cognitive maps models are merged by means of granular computing. The established model not only performs well in numerical and interval predictions but also has better interpretability. Results: Experimental studies involving both synthetic and real-life datasets demonstrate the usefulness and satisfactory efficiency of the proposed approach.

Automated summary

The study introduces a new prediction modelling methodology based on fuzzy cognitive maps, leveraging a bootstrap method to select multiple sub-sequences and merging these models through granular computing. The approach not only demonstrates strong numerical and interval prediction performance, but also improves interpretability and comprehension for humans.