A research-study submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Information Management, School of Engineering and Technology

Based on Amazon Inc.'s historical 2010-2011 data, Amazon.com new employee access forecast is based on a system designed to replace resource administrators on Amazon. Our analysis shows that the given dataset with categorical values is very unbalanced. Therefore, during the preprocessing step, we tried different sampling methods, feature selection, and differentencoding and frequency coding to make the data more suitable for prediction. In the prediction stage we first tested unique models suitable for vector machines with categorical data supportvector machine, logistic regression, light GBM and neural networks. Finally, we combine the best fourprediction results from a random forest, a gradient enhancement and a logistic regression (with encoded data) and an area under curve (AUC) improvement.