An empirical overview of nonlinearity and overfitting in machine learning using COVID-19 data

Area of Science:

• Data Science
• Epidemiology
• Machine Learning

Background:

• Accurate forecasting of infectious disease outbreaks, such as COVID-19, is crucial for public health decision-making.
• Machine learning models offer potential for predicting disease spread but require careful validation to avoid common pitfalls like overfitting.

Purpose of the Study:

• To apply support vector regression (SVR) for predicting COVID-19 cases in the 12 most affected countries.
• To evaluate the impact of different kernel functions (nonlinearity structures) on SVR model performance.
• To analyze hyperparameter sensitivity using 3-D interpolated surfaces for robust model assessment.

Main Methods:

• Support Vector Regression (SVR) was employed to forecast COVID-19 case numbers.
• Various kernel functions, including linear and Gaussian, were tested to model different levels of nonlinearity.
• Hyperparameter sensitivity was analyzed through 3-D interpolated surfaces to assess model robustness.

Main Results:

• The Gaussian Kernel, representing high nonlinearity, achieved superior in-sample fit but exhibited poor out-of-sample predictive performance, indicating overfitting.
• Conversely, the linear Kernel demonstrated weaker in-sample performance but yielded the most accurate out-of-sample forecasts.
• Model performance was sensitive to hyperparameter settings, necessitating careful tuning.

Conclusions:

• The study underscores the risk of overfitting with complex models in time-series forecasting, particularly during pandemics.
• Linear models can outperform highly nonlinear models in out-of-sample prediction for COVID-19 case forecasting.
• Findings emphasize the importance of empirical validation and cautious application of machine learning in real-world decision-making for public health crises.