Applying a Customized Ordinal Machine Learning Classification Method to Predict Potential Retail Location Performance

Presenter: Jinyuan Chen

Research Category: Computer and Information Sciences
Award Winner Category: Sharing Economy Award

Researching and opening new stores consumes considerable amounts of effort and capital. However, traditional market research techniques for location selection lack the ability to find optimal solutions with a large number of independent variables. The purpose of this project is to provide a low-cost, fast, and reliable method for identifying potential high-ROI locations with minimal human interaction.

As this project was sponsored by Puma Group, we mapped their internal ranking system to a typical ordinary variable. Their ranking uses six buckets to categorize each store based on overall store performance. These buckets amalgamate factors such as profit, customer satisfaction, expenses, etc. In addition to existing stores’ performances, we also have high-dimensionality input data regarding factors such as demographics, travel distance, tourism, socioeconomic metrics, mall brands, etc.

We used ordinal classification, a supervised ML algorithm, and trained it to map these factors onto Puma’s rankings. The custom algorithm was derived from a standard ordinal classification algorithm but was improved to avoid overfitting and deal with small sample sizes and imbalanced data. This was done by combining a series of binary classification steps in order to reconstruct the ordinal rankings. After training, the algorithm can output the probability of different ranks at potential stores.

This model can be generalized to all kinds of location selection, and it has the ability to advise retail businesses on how a potential store would be ranked. Companies can thus spend fewer resources than traditional market research and get more accurate results.