Predicting Multiple Domain Queue Waiting Time via Machine Learning

Abstract

This paper describes an implementation of the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology for a demonstrative case of human queue waiting time prediction. We collaborated with a multiple domain (e.g., bank, pharmacies) ticket management service software development company, aiming to study a Machine Learning (ML) approach to estimate queue waiting time. A large multiple domain database was analyzed, which included millions of records related with two time periods (one year, for the modeling experiments; and two year, for a deployment simulation). The data was first preprocessed (including data cleaning and feature engineering tasks) and then modeled by exploring five state-of-the-art ML regression algorithms and four input attribute selections (including newly engineered features). Furthermore, the ML approaches were compared with the estimation method currently adopted by the analyzed company. The computational experiments assumed two main validation procedures, a standard cross-validation and a Rolling Window scheme. Overall, competitive and quality results were obtained by an Automated ML (AutoML) algorithm fed with newly engineered features. Indeed, the proposed AutoML model produces a small error (from 5 to 7 minutes), while requiring a reasonable computational effort. Finally, an eXplainable Artificial Intelligence (XAI) approach was applied to a trained AutoML model, demonstrating the extraction of useful explanatory knowledge for this domain.