The Classifier Chain Generalized Maximum Entropy Model for Multi-label Choice Problems

Multi-label classification can be applied to study empirically discrete choice problems, in which each individual chooses more than one alternative. We applied the Classifier Chain (CC) method to transform the Generalized MaximumEntropy (GME) choicemodel from a single-label model to amulti-label model. The contribution of our CC-GMEmodel lies in the advantages of both the GME and CC models. Specifically, the GME model can not only predict each individual’s choice, but also robustly estimate model parameters that describe factors determining his or her choices. The CCmodel is a problem transformation method that allows the decision on each alternative to be correlated. We used Monte-Carlo simulations and occupational hazard data to compare the CC-GME model with other selected methodologies for multi-label problems using the Hamming Loss, Accuracy, Precision and Recall measures. The results confirm the robustness of GME estimates with respect to relevant parameters regardless of the true error distributions. Moreover, the CCmethod outperforms other methods, indicating that the incorporation of the information on dependence patterns among alternatives can improve prediction performance.