Recursive tree grammar autoencoders

Abstract Machine learning on trees has been mostly focused as input. Much less research investigated output, which many applications, such molecule optimization for drug discovery, or hint generation intelligent tutoring systems. In this work, we propose a novel autoencoder approach, called recursive tree grammar (RTG-AE), encodes via bottom-up parser and decodes grammar, both learned neural networks that minimize the variational loss. The resulting encoder decoder can then be utilized in subsequent tasks, time series prediction. RTG-AEs are first model to combine three features: processing, grammatical knowledge, deep learning. Our key message is unique combination of all features outperforms models any two three. Experimentally, show RTG-AE improves autoencoding error, training time, score synthetic well real datasets compared four baselines. We further prove parse generate linear expressive enough handle regular grammars.