Definitions, methods, and applications in interpretable machine learning

Abstract

Machine-learning models have demonstrated great success in learning complex patterns that enable them to make predictions about unobserved data. In addition to using models for prediction, the ability to interpret what a model has learned is receiving an increasing amount of attention. However, this increased focus has led to considerable confusion about the notion of interpretability. In particular, it is unclear how the wide array of proposed interpretation methods are related and what common concepts can be used to evaluate them. We aim to address these concerns by defining interpretability in the context of machine learning and introducing the predictive, descriptive, relevant (PDR) framework for discussing interpretations. The PDR framework provides 3 overarching desiderata for evaluation: predictive accuracy, descriptive accuracy, and relevancy, with relevancy judged relative to a human audience. Moreover, to help manage the deluge of interpretation methods, we introduce a categorization of existing techniques into model-based and post hoc categories, with subgroups including sparsity, modularity, and simulatability. To demonstrate how practitioners can use the PDR framework to evaluate and understand interpretations, we provide numerous real-world examples. These examples highlight the often underappreciated role played by human audiences in discussions of interpretability. Finally, based on our framework, we discuss limitations of existing methods and directions for future work. We hope that this work will provide a common vocabulary that will make it easier for both practitioners and researchers to discuss and choose from the full range of interpretation methods.interpretability | machine learning | explainability | relevancy

Combining Optical, Fluorescence, Thermal Satellite, and Environmental Data to Predict County-Level Maize Yield in China Using Machine Learning Approaches

“…Finally, ML and DL approaches are "black box" with limited process-based interpretation. Integrating a process-based model with data-driven approaches could not only attain interpretable ML/DL models but, more importantly, are computational efficiency and readily extrapolate outside the range of training conditions [18,91], which is recommended for future large-scale yield estimation, management optimization, and disaster monitoring.…”

Hierarchical Gradient Smoothing for Probability Estimation Trees

“…α * p = i α i and θ * p,k = i α i θ LOO i,k . Second, traverse the tree bottom-up to calculate the gradients for each internal node level by level (line [10][11][12][13][14][15][16]. Last, return the cost.…”

“…So, the problem of making a single tree perform well in inference arises, and one can ask does a single decision tree beat a random forest with 10 trees. Moreover, trees also serve as one of the few global models considered to be interpretable, an increasingly important requirement in applications [12]. Thus, quality single decision tree built efficiently have many uses.…”

Multidimensional analysis and detection of informative features in diffusion MRI measurements of human white matter

“…Though there are several 26 different available implementations of this overall idea, the principles are 27 similar [1, [21][22][23]: tractometry begins by delineating the parts of the white matter that 28 belong to different major "tracts" (i.e. anatomical or functional groups of white matter 29 fibers), such as the corticospinal tract or arcuate fasciculus, assigning tractography 30 generated streamlines to "bundles," which approximate the anatomical tracts, and 31 sampling biophysical properties (such as fractional anisotropy or mean diffusivity) along 32 the length of these bundles. the parts of the white matter that belong to different major 33 tracts (i.e.…”

“…Different 50 approaches can be taken to resolving this challenge. For example, Colby and descriptive power [29,30]. Accordingly, tractometry analysis should simultaneously 66 capitalize on all the data across all tracts to make the best possible prediction, while 67 also retaining and elucidating spatial information about the locations that are most 68 informative for a prediction.…”