In the era of Big Data, many scientific disciplines and engineering activities rely on cumulative databases, consisting of many entries derived from different experiments and studies, to investigate complex problems. Their contents can be analysed with much finer granularity than with the usual meta-analytic tools, based on summary statistics such as means and standard deviations. At the same time, not being primary studies, also traditional statistical techniques are not adequate to investigate them. New meta-analysis methods have therefore been adapted to study these cumulative databases and to ensure their validity and consistency. Information theoretic and neural computational tools represent a series of complementary techniques, which can be deployed to identify the most important variables to analyse the problem at hand, to detect whether quantities are missing and to determine the coherence between the entries provided by the individual experiments and studies. The performances of the developed methodologies are verified with a systematic series of tests with synthetic data. An application to thermonuclear fusion proves the capability of the tools to handle real data, in one of the most complex fields of modern physics.
Read full abstract