Date: March 28, 2023
Speaker: Dr. Anna Konstorum
Affiliation: Institute for Defense Analysis, Center for Computing Sciences/University of Florida
Title: Decomposition Methods For High-Dimensional Data: Making Interpretability A Priority
Abstract: As both the volume and complexity of bioinformatics data increases, it becomes pertinent to develop appropriate algorithms to transform the data into information that can be used for clustering, prediction, and evaluation of underlying biological mechanisms. The broad field of matrix and tensor decompositions, which constitute linear and multi-linear dimension reduction schemes, have been adopted by the biological community to aid in these efforts. One set of challenges that has come to light is that the original optimization schemes developed for these decompositions are not always aligned with the goal of the biologist to obtain maximum interpretability from their decompositions. We show, in the example of non-negative tensor decompositions, how this can impact the use of these models in applications. To overcome this, we develop novel evaluation metrics in order to choose models with improved interpretability properties, and discuss more generally how envisioning the decompositions as systems models (not just dimension reduction schemes), can help to improve their overall utility for the biology community. We also show the application of this philosophy to an improved implementation of a joint multi-view decomposition scheme (MCIA) that allows for expanded use and extension of MCIA for both biologists and mathematicians.