A Machine Learning Approach to Live Wait Time Estimation in a Virtual Queue App

Waiting in lines is an everyday occurrence for many people in society, and the main question people have when waiting is, ``how long will I have to wait?". With physical lines, techniques to estimate wait times require randomly sampling people to see how long they can go through the line. The data available to estimate waits has grown exponentially with the emergence of virtual queues that text customers when it is their turn in line. One of the leading companies providing virtual queue management services is Waitwhile which captured over 100 million in visits requiring waiting. Using Waitwhile data, this thesis explores machine learning to make better wait time estimations than existing mechanical-based wait time estimations. Models such as Linear Regression, XGBoost, Multi-Layer Perceptions, and Temporal Convolutional Networks were tested against four key business types (DOCTOR, HAIRDRESSER, RESTAURANT, and RETAIL). When applying a custom compression and grid search to fit the technical constraints for production, XGBoost and MLP models are on average 15 minutes more accurate than the existing rules-based model. A custom asymmetrical loss aimed to increase customer satisfaction reduced under-prediction percent for XGBoost by -12.5 ppt and MLP by -17.5 ppt. Additionally, models were trained individually by location, which performed better than generalized business type models for XGBoost (-12.3% decrease in MAE) and MLP (-37.9% decrease in MAE when using transfer-learning). When testing a time-series method, TCN did not consistently outperform XGBoost or MLP. XGBoost was the most consistent performing model that often produced the lowest MAE across every experiment. The final implementation of XGBoost for production was developed using a custom-built method in Node.js and fell within the necessary latency and memory requirements for production in Waitwhile.