Learning in the Presence of Large Fluctuations: A Study of Aggregation and Correlation

Consider a scenario where one aims to learn models from data being characterized by very large fluctuations that are neither attributable to noise nor outliers. This may be the case, for instance, when predicting the potential future damages of earthquakes or oil spills, or when conducting financial data analysis. If follows that, in such a situation, the standard central limit theorem does not apply, since the associated Gaussian distribution exponentially suppresses large fluctuations. In this paper, we present an analysis of data aggregation and correlation in such scenarios. To this end, we introduce the Levy, or stable, distribution which is a generalization of the Gaussian distribution. Our theoretical conclusions are illustrated with various simulations, as well as against a benchmarking financial database. We show which specific strategies should be adopted for aggregation, depending on the stability exponent of the Levy distribution. Our results indicate that the correlation in between two attributes may be underestimated if a Gaussian distribution is erroneously assumed. Secondly, we show that, in the scenario where we aim to learn a set of rules to estimate the level of stability of a stock market, the Levy distribution produces superior results. Thirdly, we illustrate that, in a multi-relational database mining setting, aggregation using average values may be highly unsuitable.