Ensembles of Interesting Subgroups for Discovering High Potential Employees
Published on Apr 19, 2016 in KDD (Knowledge Discovery and Data Mining)
· DOI :10.1007/978-3-319-31750-2_17
We propose a new method for building a classifier ensemble, based on subgroup discovery techniques in data mining. We apply subgroup discovery techniques to a labeled training dataset to discover interesting subsets, characterized by a conjuctive logical expression rule, where such subset has an unusually high dominance of one class. Treating these rules as base classifiers, we propose several simple ensemble methods to construct a single classifier. Another novel aspect of the paper is that it applies these ensemble methods, along with standard anomaly detection and classification, to automatically identify high potential HIPO employees - an important problem in management. HIPO employees are critical for future-proofing the organization in the face of attrition, economic uncertainties and business challenges. Current HR processes for HIPO identification are manual and suffer from subjectivity, bias and disagreements. Proposed data-driven analytics algorithms address some of these issues. We show that the new ensemble methods perform better than other methods, including other ensemble methods on a real-life case-study dataset of a large multinational IT services company.