totalvis: A Principal Components Approach to Visualizing Total Effects in Black Box Models

While a wide variety of machine-learning techniques have been productively applied to diverse prediction tasks, characterizing the nature of patterns and rules learned by these techniques has remained a difficult problem. Often called ‘black-box’ models for this reason, visualization has become a prominent area of research in understanding their behavior. One powerful tool for summarizing complex models, partial dependence plots (PDPs), offers a low-dimensional graphical interpretation by evaluating the effect of modifying individual predictors on fitted/predicted values. Nevertheless, in high-dimensional settings, PDPs may not capture more complex associations between groups of related variables and the outcome of interest. We propose an extension of PDPs based on the idea of grouping covariates, and interpreting the total effects of the groups. The method utilizes principal components analysis to explore the structure of the covariates, and offers several plots for assessing the approximation function. In conjunction with our diagnostic plot, totalvis gives insight into the total effect a group of covariates has on the prediction and can be used in situations where PDPs may not be appropriate. These tools provide a useful approach for pattern exploration, as well as a natural mechanism to reason about potential causal effects embedded in black-box models.