Supplier classification by applying AutoML

Abstract

The interest of this research referred to the possibility of detecting patterns to identify the best supplier in a citrus exporter under a classification model that would provide a high degree of confidence. This work aims to verify the validity of Automated Machine Learning (AutoML) employing automatic model selection and hyperparameter optimization method applied in suppliers classification. These results can be used as support for decision-makers to evaluate and select the best of their alternatives. 786 historical records of suppliers were analyzed. The historical records were separated into datasets classified by season (high, low) and production area (north, center, and south). The sampling criteria implemented on the datasets were: cross-validation, percentage split (66%), and representative sample. The classification was evaluated by employing confusion matrix and performance indicators for each dataset according to the sampling criterion chosen. Through AutoML, the following algorithms: Vote, Random Forest, Attribute-Selected-Classifier, Bayes Network, Repeated Incremental Pruning to Produce Error Reduction (RIPPER), and Naive Bayes found the best percentage of sensitivity (89.44, 70.79, 87.50, 88.59, 76.62, 90.13) for each dataset respectively under the highest performance sampling criterion. The AutoML approach helped discriminate Persian lemon suppliers based on their sourcing history, providing support for decision-making under acceptable percentages in reliability indicators.