Comparison of Neural Networks and Gradient Boosting Models on Ordinal Age Class Prediction using Mouse Trajectories

Tobias Wistuba, Lisa Bondo Andersen, Ailin Liu, Felix Henninger, Frauke Kreuter, Sonja Greven

Keywords: deep learning, mouse trajectories, ordinal regression, classification, boosting

Abstract: We aim to predict ordinal age classes via mouse trajectories collected through web surveys. We compare performance of different neural network architectures, including dense neural networks using the entire trajectory as input, 1D- and 2D-convolutional neural networks, long short-term memory neural networks, and transformers against gradient boosting models that use hand-crafted features of the trajectories as inputs. The results show that neural networks as well as gradient boosting models are able to predict age classes with accuracies above pure chance. However, despite their higher complexity, neural networks do not clearly outperform the boosting models and do not offer the advantages of supplying interpretable results and detecting informative covariates.

Serve As Reviewer: ~Tobias_Wistuba1

Submission Number: 38