PRISM: A Rich Class of Parameterized Submodular Information Measures for Guided Data Subset Selection

With ever-increasing dataset sizes, subset selection techniques are becoming increasingly important for a plethora of tasks. It is often necessary to guide the achieve certain desiderata, which includes focusing or targeting data points, while avoiding others. Examples such problems include: i)targeted learning, where goal find subsets with rare classes attributes on model under performing, and ii)guided summarization, (e.g., image collection, text, document video) summarized quicker human consumption specific additional user intent. Motivated by applications, we present PRISM, rich class PaRameterIzed Submodular information Measures. Through novel functions their parameterizations, PRISM offers variety modeling capabilities that enable trade-off between desired qualities like diversity representation similarity/dissimilarity set points. We demonstrate how can be applied two real-world mentioned above, require guided selection. In doing so, show interestingly generalizes some past work, therein reinforcing its broad utility. extensive experiments diverse datasets, superiority over state-of-the-art in targeted learning image-collection summarization. available as part SUBMODLIB (https://github.com/decile-team/submodlib) TRUST (https://github.com/decile-team/trust) toolkits.