ExTuNe: Explaining Tuple Non-conformance

2020 International Conference on Management of Data |

Published by ACM

Download PDF | View Publication

In data-driven systems, we often encounter tuples on which the predictions of a machine-learned model are untrustworthy. A key cause of such untrustworthiness is non-conformance of a new tuple with respect to the training dataset. To check conformance, we introduce a novel concept of data invariant, which captures a set of implicit constraints that all tuples of a dataset satisfy: a test tuple is non-conforming if it violates the data invariants. Data invariants model complex relationships among multiple attributes; but do not provide interpretable explanations of non-conformance. We present ExTuNe, a system for Explaining causes of Tuple Non-conformance. Based on the principles of causality, ExTuNe assigns responsibility to the attributes for causing non-conformance. The key idea is to observe change in invariant violation under intervention on attribute-values. Through a simple interface, ExTuNe produces a ranked list of the test tuples based on their degree of non-conformance and visualizes tuple-level attribute responsibility for non-conformance through heat maps. ExTuNe further visualizes attribute responsibility, aggregated over the test tuples. We demonstrate how ExTuNe can detect and explain tuple non-conformance and assist the users to make careful decisions towards achieving trusted machine learning.

ExTuNe: Explaining Tuple Non-conformance

Authors: Anna Fariha, Ashish Tiwari, Arjun Radhakrishna, Sumit Gulwani Abstract: In data-driven systems, we often encounter tuples on which the predictions of a machine-learned model are untrustworthy. A key cause of such untrustworthiness is non-conformance of a new tuple with respect to the training dataset. To check conformance, we introduce a novel concept of data invariant, which captures a set of implicit constraints that all tuples of a dataset satisfy: a test tuple is non-conforming if it violates the data invariants. Data invariants model complex relationships among multiple attributes; but do not provide interpretable explanations of non-conformance. We present ExTuNe, a system for Explaining causes of Tuple Non-conformance. Based on the principles of causality, ExTuNe assigns responsibility to the attributes for causing non-conformance. The key idea is to observe change in invariant violation under intervention on attribute-values. Through a simple interface, ExTuNe produces a ranked list of the test tuples based on their degree of non-conformance and visualizes tuple-level attribute responsibility for non-conformance through heat maps. ExTuNe further visualizes attribute responsibility, aggregated over the test tuples. We demonstrate how ExTuNe can detect and explain tuple non-conformance and assist the users to make careful decisions towards achieving trusted machine learning. Paper link: https://dl.acm.org/doi/abs/10.1145/3318464.3384694