On Efficient Approximate Queries over Machine Learning Models

The question of answering queries over ML predictions has been gaining attention in the database community. This is challenging because finding high quality answers by invoking an oracle such as a human expert or expensive deep neural network model on every single item DB and then applying query, can be prohibitive. We develop novel unified framework for approximate query leveraging proxy to minimize usage both Precision-Target (PT) Recall-Target (RT) queries. Our uses judicious combination data samples cheap objects. It relies two assumptions. Under P roxy Q uality assumption, we algorithms: PQA that efficiently finds with probability no calls, PQE, heuristic extension achieves empirically good performance small number calls. Alternatively, under C ore S et losure CSC returns minimal usage, CSE, which extends it more general settings. extensive experiments five real-world datasets types, PT RT, demonstrate our algorithms outperform state-of-the-art achieve result provable statistical guarantees.