Abstract
Choosing input variables and network architecture are key processes for modeling short-term incidence rate forecasts using artificial neural networks. This paper proposes a method based on Rough Set Theory (RST) to address these challenges. In the proposed approach, key factors affecting incidence rate forecasting are first identified by RST to determine the model's input variables. Subsequently, a set of influence rules is obtained through the reductive mining of attributes and attribute values. A neural network incidence rate forecast model is then established based on this rule set, and the Backpropagation (BP) algorithm is adopted to optimize the network. This method indicates that incidence rate forecast models can be established according to theoretical principles, avoiding the "blindness" of traditional trial-and-error approaches. A practical application is provided to demonstrate the utility of this novel method.
Introduction
Artificial Neural Networks (ANNs) are highly effective at predicting non-linear trends, such as disease incidence rates. However, the design process for these networks is often criticized for being "blind," relying heavily on researcher intuition or trial and error to determine input variables and hidden layer architectures. This lack of formal methodology often results in redundant data processing, increased noise, and overly complex models that are prone to overfitting and lack interpretability.
This paper introduces a hybrid framework that integrates Rough Set Theory (RST) with Backpropagation (BP) Neural Networks. By utilizing RST to mathematically prune redundant inputs and define the network's internal structure through logical rules, the authors transform the ANN from a "black box" into a theoretically grounded, transparent forecasting tool.
Methodology
The proposed pipeline utilizes Rough Set Theory as both a pre-processor for data and an architectural blueprint for the neural network. The process follows four distinct stages:
- Discretization: Since RST operates on discrete values, continuous meteorological and incidence data are partitioned into specific intervals or regions.
- Attribute Reduction: Using RST "decision tables," the authors identify the attributes that significantly influence the outcome. By calculating the "importance" of each attribute, redundant factors that do not contribute to the table's compatibility degree are discarded.
- Rule Extraction: Reasoning rules (if-then logic) are mined from the reduced data set.
- ANN Construction: The final set of reduced attributes defines the input layer, while the number of hidden neurons is set to match the number of extracted reasoning rules. The network is then optimized using a standard BP algorithm with momentum.
Experimental Setup and Datasets
The model was validated using a real-world dataset from Qinghai Haixizhou First People's Hospital in China, covering the period from January 2003 to December 2009. The variables included monthly maximum incidence rates, average temperature, average monthly air pressure, and average humidity. Data from 2003–2008 was utilized for model training, while the 2009 data served as the evaluation set.
Key Results and Discussion
The RST-ANN approach demonstrated significant improvements in both efficiency and accuracy. Through attribute reduction, the initial pool of 12 potential input variables was condensed to just 4 key factors, simplifying the model without compromising predictive power.
The mapping of logical rules directly to hidden neurons provided a mathematical justification for the network's size, enhancing explicability. In testing, the model achieved an average relative error of 0.925% after 1,000 learning cycles, with a remarkably fast training time of 16ms.
When compared to traditional alternatives like the Sugeno fuzzy forecasting model, the RST-ANN model required fewer input variables and rules while achieving higher precision (the Sugeno model typically yields error rates between 1.4% and 1.6%).
Conclusion
The integration of Rough Set Theory into the ANN design process effectively mitigates the "blindness" of traditional neural network modeling. While the approach offers high accuracy and efficiency, the authors note that output space partitioning remains a challenge; overly fine partitioning can lead to a "rule explosion" that complicates the network. Future work will focus on optimizing the division of the output space to maintain high accuracy while minimizing model complexity.
Summary updated