The Effects of Different Parameters on the Accuracy of Deep Learning Models for Predicting U.S. Citizen’s Life Expectancy

Abstract

The increasing prevalence of deep learning-based machines in the daily life of average people results in a demand for research to be done on improving the accuracy of deep learning models. In response to this need, this paper aims to explore the effects of changing the parameters of a deep learning model, including the neuron count, epoch count, batch size, and validation split on the prediction accuracy of a deep learning model. We used the programming language Python, the TensorFlow and Pandas libraries, and the Keras application programming interface to create 13 regression-based deep learning models, all but one, which was used as a standard, of which had a parameter altered to be lower or higher than the standard model. After training each model using a dataset comprised of data from the 2010 United States census, we measured the predictive accuracy of each model at different epoch counts using the absolute average difference between the predictions of life expectancy from the models and the actual value from the 2010 U.S. census dataset. Based on the absolute average difference for each model, we found that increasing the neuron count, epoch count, and batch size and decreasing the validation split improves prediction accuracy in deep learning models, in most cases. These results can be used to create more accurate deep learning models for scientific or commercial use, and the models themselves can be used for their ability to predict life expectancies from given data, based on learned trends.