A Bump Hunting Approach to Finding Interpretable Data Pockets

This dissertation approaches the problem of extracting simple interpretations from local regions of data. This is sometimes called bump hunting because the local regions of interest have a high concentration of a particular output value. This work develops a bump hunting method for discrete-valued tabular data where each bump is modeled by a rectangular region of the input data space so its rule-based description admits a simple logical interpretation that can inform decisions. This method is designed for labeled data where each input feature has a distinct meaning that may or may not be related to the output, and the goal is to find feature subsets related to the output, rectangular regions within these subset feature spaces, and pockets of data within these rectangular regions that simultaneously obey five properties: each rectangle is described by a small subset of input features, the pocket data occupies a local region of the subset-feature space, the input/output relationship for the pocket data is nearly pure, the number of pocket data samples in each rectangle is both statistically significant and large enough to be relevant for the end application, and the overlap between rectangles is minimal. In contrast to state-of-the-art methods that use decision trees or the PRIM algorithm, this new method is better at distinguishing closely spaced bumps, representing non-rectangular shaped bumps formed by co-linear features, and controlling the extent of the rectangles, and more robust against overfitting and inclusion of spurious features with little or no relation to the output.