Bayesian Classification of Personal Histories - An application to the Obesity Epidemic

Abstract

Bayesian classifiers are an important tool in machine learning. The Naive Bayes classifier in particular has shown itself to be a robust, computationally efficient and transparent approximation. However, due to its strong assumption of feature independence it is often too easily discarded. Understanding under what circumstances feature correlations are likely to occur and how to diagnose them is an important first step in improving the Naive Bayes approximation. Furthermore, combining features can lead to enhanced approximations. In this paper we show the benefits of a Generalised Bayes Approximation that accounts for feature correlations while, at the same time, showing that an important area where feature correlations are ubiquitous is in time series that represent aspects of human behaviour. Using data that represent historical patterns of exercise we show how more predictive and more insightful models for obesity can be constructed using a Generalised Bayes approximation that combines historical features, thereby capturing the idea of human habits. In particular, by analysing feature combinations we show that abandoning good past exercise habits is more correlated with obesity than never having had them in the first place.