Quantum Machine Learning Applied to the Sinking of the Titanic

Abstract

This work investigates the performance of hybrid quantum-classical variational classifiers applied to a supervised learning task involving the titanic3 dataset. Quantum models were constructed using Pauli entangling and non-entangling expansion-based feature maps and the RealAmplitudes ansatz with up to 50 variational parameters. Model training employed the COBYLA gradient-free optimizer to minimize the cross-entropy loss, within an ideal statevector simulation framework. Comparative performance analysis reveals that the models based on the non-entangling feature map consistently outperformed the models based on the entangling features maps, achieving saturation of classification metrics (accuracy, balanced accuracy, and Youden's index) beyond 15 to 20 parameters. Further, two quantum models were benchmarked against a classical Support Vector Classifier (SVC). While both approaches yielded similar predictive performance across multiple training sizes, the classical model exhibited a performance collapse when trained with 90% of the dataset, a failure mode absent in the quantum classifiers. These results underscore the robustness and viability of variational quantum classifiers for binary classification tasks on classical datasets in the NISQ era.