Retail Store Risk Prediction from Imbalanced Data

Abstract

In the literature, the domain of risk prediction has received considerable attention. However, the particular area of retail store risk prediction has been addressed by only a few researchers, with most of the previous work focused on two domains: Public Health and Finance. This motivates the need for conducting a study that addresses this area. In this project, we rethink the retail store risk prediction problem by mining a dataset that contains store license applications, these applications represent the stores in Jeddah city. Similar to previous work, the problem is formulated as a binary classification task. The prediction features are identified and the class label is extracted. Conventional models are the most popular in the domain of risk prediction, therefore our methodology involves designing and fine-tuning ten different conventional classifiers that are used to build an ensemble. The dynamic ensemble selection approach is adopted rather than the static ensemble. In real life, and in risk prediction problems, datasets suffer from class imbalanced issues; thus 4 resampling methods are considered to further enhance the performance. Moreover, a RESTful store risk prediction web service that follows the OAuth V2 protocol is design to be easily and securely consumed by the public, with the developed ensemble operating as the backend of this service. The service is successfully deployed and tested. It achieved the project main objectives, in addition, it obtained good generalisation performance, outperforming all the ten conventional models.