Extensible TensorFlow Implementations of Projection Quality Metrics

Dimensionality Reduction (DR, also called Projection) algorithms enable the exploration of high-dimensional data by generating low-dimensional representations of it - typically 2D or 3D scatterplots. Such representations are designed to map data patterns to visual patterns analyzable by humans. Projections can vary wildly - even for a fixed dataset - depending on technique and hyperparameters chosen and, as such, do not all preserve all data patterns equally well. To assess this, so-called Projection Quality Metrics (PQMs) are used. However, the ever-growing number of Projection Quality Metrics has led to fragmented implementations which hinder their easy reuse, leading in turn to unequal adoption and inconsistent implementations. In this work, we propose a TensorFlow-based library of PQMs, improving the previous state of the art in terms of ergonomics, extensibility, and computational scalability. We discuss our improvements and elicit areas where the gap between implementation and research is significant in the area of Projection Quality Metrics, pointing to avenues for future work in developing better PQM libraries that aim to fill this gap.