Explaining inference queries with bayesian optimization

Obtaining an explanation for SQL query result can enrich the analysis experience, reveal data errors, and provide deeper insight into data. Inference seeks to explain unexpected aggregate results on inference data; such queries are challenging because may need be derived from source, training, or in ML pipeline. In this paper, we model objective function as a black-box propose BOExplain, novel framework explaining using Bayesian optimization (BO). An is predicate defining input tuples that should removed so of interest significantly affected. BO --- technique finding global optimum used find best predicate. We develop two new techniques (individual contribution encoding warm start) handle categorical variables. perform experiments showing predicates found by BOExplain have higher degree compared those state-of-the-art engines. also show effective at deriving explanations source training variety real-world datasets. open-sourced Python package https://github.com/sfu-db/BOExplain.