All Times ET
Keywords: Designed data, unstructured data, predictive modeling, super large parameter problems
In this presentation, we will highlight Statisticians missed opportunities in making significant contributions to big data analytics. The presentation will explore building interpretable predictive models using a smaller number of parameters and known well-established statistical principles, along with the geometrical and dimensional structures of unstructured data. Statisticians have been successful in solving problems using optimality principles with limited number of parameters for designed and structured data. With explosion of growth in computer hardware, software, and Big-data, the need for “huge” parameter models with unstructured data has become a necessity for prediction. Computer scientists have significant innovations as machine learning/deep learning engineering for such unstructured data scenarios. For example, image, voice, and text analytics are currently performed with a huge number of parameters using “black box” machine learning algorithms, neural networks, and artificial intelligence. While interpretable machine learning methods are gaining momentum, statisticians have a transformative power if they engage in integrating age-tested statistical designs, structures, and measures with interpretable machine learning methods with structured computations rather than “perceptron” computation. Such models can provide structured thinking in models of prediction for “huge” parameter unstructured data opportunities, beyond “perceptron” computations, which is harder to interpret and harder to believe beyond the claim that it resembles humanistic neuron type relationships among computational quantities. This redirects statisticians to model “millions of parameters” problems using proven statistical measures of centrality, dispersion, association, clustering, geometric skewing, rotation, reflection, dimensional modeling, and causality.