A Robust MCMC-based Method for Piecemeal Estimation of Distributional Features in Continuous Data
Raed A. Said · International Journal of Applied Mathematics & Statistics/International journal of applied mathematics and statistics · 2015
In the big data era data mining applications typically rely on learning algorithms to extract subtle knowledge from data attributes. The veracity of such algorithms is a function of training, validation and test data - usually sampled from high-dimensional multi-faceted data. Yet, despite the interests and the analytical tools available, attaining data modelling veracity in the big data era remains a huge challenge mainly due to the dynamics in data volumes and varieties. One commonly used method of estimation is the Markov Chain Monte Carlo (MCMC) simulation which involves drawing large random samples from a known probability distribution. Its main idea is that as the number of samples grows, the estimate converges to the true expectation parameter. The paper proposes an MCMC-based method for sampling from high-dimensional multi-faceted data. Its main idea is to discretise any given data vector at different points irrespective of known class boundaries – thus, yielding different density estimates when are then compared for accuracy. Implementation on 65436 data points (861 seismic signals of 76 observations each) and 3160 open-source monthly average sunspots readings exhibited unprecedented robustness.