Casual inference and the small data problem.
Causal inference goes one critical step beyond classical statistical inference, which limits itself to inferring the likelihood of a correlation, difference or combined effect size among observations. Causal inference synthesizes the combined likelihood of a series of statistical tests in order to establish causality. The relative increase in the number of hypotheses tested on the same data requires a relative increase in the number of samples collected to reach a comparable inference power (as in the model selection problem). In actual practice, as opposed to simulations, data sizes are often critically limited. Common reasons include paucity of historically archived data, sparse sensor deployment, limited recording time available or prohibitive incremental cost of collecting samples (often the case in neural recordings and clinical tests). Additionally, even ‘big data’ sized time series recording may contain fundamental nonstationarities which effectively segment it into numerous ‘small data’ regions. Examples from different domains (human data, energy consumption) will be shown to highlight relevant challenges and need for strong priors in ‘small data’ contexts.
BIO:Dr. Florin Popescu graduated from Purdue University, USA in 1989. He received his Ph.D. in Mechanical Engineering from Northwestern University, USA in 2000, and is a Fraunhofer researcher since 2005. He has received awards such as the Marie Curie Excellence Team grant (2005-2009) for young researchers, having led a multi-disciplinary team of researchers in multimodal sensor-assisted robotics and led several other projects at Fraunhofer in renewables forecasting, machine learning approaches to industrial energy management and efficient deployment of smart sensors in urban environments. Among other research interests are automated diagnostics and causal inference in dynamic systems. He is the author of over 30 peer-reviewed articles, including PlosONE, proc. ICAR and NeuroImage and has co-edited a book on Causality in Time Series Research (JMLR W&CP series).