Recurrent neural networks for dynamic user intent prediction in human-database interaction

Prediction of human intent during a user interaction session with the database received a significant amount of attention in the recent past [1, 8]. State-of-the-art intent detection approaches such as [5] insist that the human intent is dynamic and is constantly changing throughout the user session. While the usage of classifiers like SVMs and decision trees have been proposed to capture static user intent [2], such models become ineffective in predicting dynamic or ever-changing human intent. Recurrent Neural Networks (RNNs) are powerful temporal predictors and have recently been prominent in the database research community for tasks such as entity matching [3, 6]. In this work, we discuss the application of RNNs to the problem of dynamic user intent prediction during a human-database interaction. We propose two variants of SQL-specific embedding vectors for RNNs. We also propose active learning strategies for RNNs which consume a fraction of the held-out training data to produce competitive prediction quality as full training or supervised learning. Our experiments on real user sessions upon the NYCTaxiTrip dataset [9] evaluate the effectiveness of vanilla, LSTM and GRU based RNNs.