Partially autoregressive machine learning: Development and testing of methods to predict United States Air Force retention
Document Type
Article
Publication Date
9-2022
Abstract
Establishing effective personnel management policies in the United States Air Force (USAF) requires methods to predict the number of personnel remaining in the USAF for different lengths of time in the future. Defined as the Personnel Retention Problem (PRP), determining this type of aggregate survival rate is a time series regression problem that shares many characteristics with binary classification problems. The limitations of this particular structure are particularly difficult to overcome for problems with limited data like the USAF PRP. We develop and test several machine learning models to produce improved retention predictions compared to the USAF’s current Kaplan Meier model. In addition to traditional random forest models and feedforward neural networks, we propose the inclusion of a partially autoregressive feature to extend the benefits of low-capacity autoregressive techniques to higher-capacity machine learning techniques. We present a Partially Autoregressive Neural Network (PARNet) and a Partially Autoregressive Random Forest (PARFor) and test the performance of each technique across a range of hyperparameter values. We select the superlative model using a validation dataset, compare results to the existing benchmark model, and find a 62.8% reduction in aggregate prediction error for the baseline neural network and 34.8% reduction for the PARNet.
Source Publication
Computers & Industrial Engineering (ISSN 0360-8352 | eISSN 1879-0550)
Recommended Citation
Hoecherl, J. C., Robbins, M. J., Borghetti, B. J., & Hill, R. R. (2022). Partially autoregressive machine learning: Development and testing of methods to predict United States Air Force retention. Computers & Industrial Engineering, 171, Art. 108424. https://doi.org/10.1016/j.cie.2022.108424
Comments
The full text of the published article is available via subscription or purchase at Elsevier's ScienceDirect portal, through the DOI link below.