Adversarial Autoencoder and Multi-Task Semi-Supervised Learning for Multi-stage Process

Abstract

In selection processes, decisions follow a sequence of stages. Early stages have more applicants and general information, while later stages have fewer applicants but specific data. This is represented by a dual funnel structure, in which the sample size decreases from one stage to the other while the information increases. Training classifiers for this case is challenging. In the early stages, the information may not contain distinct patterns to learn, causing underfitting. In later stages, applicants have been filtered out and the small sample can cause overfitting. We redesign the multi-stage problem to address both cases by combining adversarial autoencoders (AAE) and multi-task semi-supervised learning (MTSSL) to train an end-to-end neural network for all stages together. The AAE learns the representation of the data and performs data imputation in missing values. The generated dataset is fed to an MTSSL mechanism that trains all stages together, encouraging related tasks to contribute to each other using a temporal regularization structure. Using real-world data, we show that our approach outperforms other state-ofthe-art methods with a gain of 4x over the standard case and a 12% improvement over the second-best method.