Interpretable Feature Generation using Deep Neural Networks and its Application to Lane Change Detection

This paper proposes a novel approach towards generating interpretable features from time series using deep neural networks. It is shown that by exploiting the advantages of heatmapping techniques, the dimensions of an interpretable domain can be strongly reduced while introducing new and relevant features to the dataset. It is then possible to learn interpretable classifiers with these features. The proposed method is a combined feature generation and selection approach that increases accuracy as compared to reference techniques while remaining fully interpretable. The methodology uses the fact that the ouput of the first convolutional layer can preserve interpretability in time series data. The application of classifying lane change intentions is considered and the public dataset highD is used to validate the approach. The results show that the generated features yield better classification results than compared reference feature sets, thereby proving the high relevance of the features that were generated with the help of deep neural networks. Especially the interpretable binary decision tree classifier shows a strong improvement in accuracy when trained on the generated feature set, providing a powerful and from input to output interpretable classification method in combination.