Divisi: Interactive Search and Visualization for Scalable Exploratory Subgroup Analysis

Re-rank subgroups by criteria such as rate and coverage of an outcome B Visualize subgroup overlap in a map of the dataset E Selected subgroups are designated with consistent colors throughout interface Ranking functions and metrics can be defined in Python code or interactively Interface embedded in a computational notebook for frictionless setup Compare subgroup metrics at a glance

C Edit subgroup definitions to see how alternative values affect metrics D Run algorithm to find data subgroups with interesting differences A Save important subgroups for further review F Figure 1: Divisi is an interactive visualization system to help data scientists perform exploratory subgroup analysis on large datasets with many feature dimensions, such as the dataset of airline passenger satisfaction ratings shown [1]. Implemented as a computational notebook widget, Divisi includes a novel approximate subgroup discovery algorithm (A) which allows interactive re-ranking by customizable functions, such as error rate and coverage (B). Users can compare metrics across subgroups (C) and test alternative rule definitions (D) to evaluate subgroups. Finally, the Subgroup Map (E) depicts overlap and coverage between groups, so users can curate the most representative subgroups for review (F).