A0534
Title: Predictive modeling for coal production locations with kernel density and some machine learning approaches
Authors: Karthik Sriram - Indian Institute of Management India (India) [presenting]
Abstract: For a major coal mining company in India, the problem of predicting which of its mining locations will produce transportable coal on any given day is addressed. The performance of different predictive machine learning approaches is adapted and evaluated, including random forests, support vector machines, logistic regression, linear discriminant analysis, multilayer perceptrons, and the long short-term memory model. In addition, motivated by the context and structure of the coal production data, a novel approach is devised based on kernel density estimation (KDE) with adaptive bandwidths, by accounting for the discrete spatial locations of the mines as well as their varying frequencies of production. The different modeling approaches are evaluated using standard predictive accuracy measures for classification problems, based on recent production data from 136 different mines (owned by the coal mining company) over one year. The results suggest that the KDE-based approach outperforms standard machine learning approaches, highlighting the importance of devising novel modeling solutions motivated by the structure of contextual data.
