Professional employment agencies often collect background data and run various personality and professional tests for candidates when considering job placements. At the same time, the final evaluation of the candidate is usually subjective and relies on the impression of the recruiter and on his specific opinion. Despite those efforts, it is reported that there is a low validity of personality tests for predicting job performance, while the accuracy given in the literature stands around fifteen percent (Morgeson, et al., 2007).
In order to improve those results, while affected by the increasing availability of HR data, there is a growing use of machine-learning algorithms to aid with the prediction task of a successful recruitment. A key feature in the required machine learning models for this task is their interpretability, such that the recruiter can gain insights regarding the candidate and the person-job fit during the evaluation process. Otherwise, the recruiter can find it difficult to accept or reject a candidate based on a ‘black box’ algorithm and without some explainable recommendations to take it into account.
An additional challenge for such a model would be its ability to find the relevant person/job features out of a rich and wide set of HR attributes. The required model should also perform well on categorical features with many different values (such as place of birth, academic background, cultural background and role) as well as features with missing values.
Decision trees are a well-known family of algorithms that address the above-mentioned requirements, however they often suffer from a big variance and overfit to the training dataset, resulting in poor generalization and prediction ability on a test set. Therefore, in this study we propose to apply a Variable
Order Bayesian Networks model (VOBN), as proposed by Ben-Gal et al., (2005) and Singer and Ben-Gal (2007), to address these HR analytics tasks. This VOBN model is not only interpretable but also flexible enough for mining significant patterns that can be used for classification and prediction of successful or unsuccessful recruitments and placements.
Our unique dataset includes hundreds of thousands of employees that were recruited to a variety of jobs in a very large organization. The data was collected over 10 years, and contains 162 features, including labeled ones by the HR personnel of the organization.
We use this unique dataset to compare the performance of the proposed model with a set of interpretable models, in terms of precision, recall and the generalization ability. We also reveal interesting insights, that are based on extracted patterns of features that can be used to predict successful assignments. The patterns are sometimes counter intuitive and shed a light on the limitations of existing theories to predict or explain successful employees’ recruitments. These findings can contribute to a more strategic organizational recruitment process as discussed in the talk.
Keywords: Recruitment, Turnover, Machine Learning, Human Resource Analytics