Partial dependence plots (PDPs) and individual conditional expectation (ICE) plots are visualizations used for explaining the behavior of machine learning (ML) models trained on tabular datasets. They show how the values of a feature or pair of features impact a model's predictions. However, in models with a large number of features, it is impractical for an ML practitioner to analyze all possible plots. To address this, we present new techniques for ranking and filtering PDP and ICE plots and build upon existing strategies for clustering the lines in ICE plots. Together, these techniques aim to help ML practitioners efficiently explore PDP and ICE plots and identify interesting model behavior. We integrate these techniques into PDPilot, a visual analytics tool that runs in Jupyter notebooks. We use PDPilot to study how 7 ML practitioners utilize the ranking, filtering, and clustering techniques to analyze an ML model.

The emergence of distinct machine learning explanation methods has leveraged a number of new issues to be investigated. The disagreement problem is one such issue, as there may be scenarios where the output of different explanation methods disagree with each other. Although understanding how often, when, and where explanation methods agree or disagree is important to increase confidence in the explanations, few works have been dedicated to investigating such a problem. In this work, we proposed Visagreement, a visualization tool designed to assist practitioners in investigating the disagreement problem. Visagreement builds upon metrics to quantitatively compare and evaluate explanations, enabling visual resources to uncover where and why methods mostly agree or disagree. The tool is tailored for tabular data with binary classification and focuses on local feature importance methods. In the provided use cases, Visagreement turned out to be effective in revealing, among other phenomena, how disagreements relate to the quality of the explanations and machine learning model accuracy, thus assisting users in deciding where and when to trust explanations. To assess the effectiveness and practical utility of Visagreement, we conducted an evaluation involving four experts. These experts assessed the tool's Effectiveness, Usability, and Impact on Decision-Making. The experts confirm the Visagreement tool's effectiveness and user-friendliness, making it a valuable asset for analyzing and exploring (dis)agreements.

Deep neural networks have revolutionized many fields, but their black-box nature also occasionally prevents their wider adoption in fields such as healthcare and finance where interpretable and explainable models are required. The recent development of Neural Additive Models (NAMs) poses a major step in the direction of interpretable deep learning for tabular datasets. In this paper, we propose a new subclass of NAMs that utilize a single-layer neural network construction of the Gaussian process via random Fourier features, which we call Gaussian Process Neural Additive Models (GP-NAM). GP-NAMs have the advantage of a convex objective function and number of trainable parameters that grows linearly with feature dimensions. It suffers no loss in performance compared with deeper NAM approaches because GPs are well-suited to learning complex non-parametric univariate functions. We demonstrate the performance of GP-NAM on several tabular datasets, showing that it achieves comparable performance in both classification and regression tasks with a massive reduction in the number of parameters.

We propose a Gaussian process approach to collaborative filtering based on the random Fourier feature method. The result is a neural network consisting of two non-linear layers in which only the weights of the middle layer need to be learned, along with the embeddings of each user and item. The model is asymptotically equivalent to a Gaussian process with RBF kernel on the space of concatenated user/item embeddings. It has the advantage of computational scalability and doesn't require defining inducing points required by other Gaussian process approximation methods. The compactness of our representation and interpretability our model can make it attractive to a variety of practitioners in applied settings. We illustrate our model on a political voting dataset, followed by more extensive experiments on several MovieLens datasets of different sizes.

With the increasing use of black-box Machine Learning (ML) techniques in critical applications, there is a growing demand for methods that can provide transparency and accountability for model predictions. As a result, a large number of local explainability methods for black-box models have been developed and popularized. However, machine learning explanations are still hard to evaluate and compare due to the high dimensionality, heterogeneous representations, varying scales, and stochastic nature of some of these methods. Topological Data Analysis (TDA) can be an effective method in this domain since it can be used to transform attributions into uniform graph representations, providing a common ground for comparison across different explanation methods. We present a novel topology-driven visual analytics tool, Mountaineer, that allows ML practitioners to interactively analyze and compare these representations by linking the topological graphs back to the original data distribution, model predictions, and feature attributions. Mountaineer facilitates rapid and iterative exploration of ML explanations, enabling experts to gain deeper insights into the explanation techniques, understand the underlying data distributions, and thus reach well-founded conclusions about model behavior. Furthermore, we demonstrate the utility of Mountaineer through two case studies using real-world data. In the first, we show how Mountaineer enabled us to compare black-box ML explanations and discern regions of and causes of disagreements between different explanations. In the second, we demonstrate how the tool can be used to compare and understand ML models themselves. Finally, we conducted interviews with three industry experts to help us evaluate our work.

One of the potential solutions for model interpretation is to train a surrogate model: a more transparent model that approximates the behavior of the model to be explained. Typically, classification rules or decision trees are used due to their logic-based expressions. However, decision trees can grow too deep, and rule sets can become too large to approximate a complex model. Unlike paths on a decision tree that must share ancestor nodes (conditions), rules are more flexible. However, the unstructured visual representation of rules makes it hard to make inferences across rules. In this paper, we focus on tabular data and present novel algorithmic and interactive solutions to address these issues. First, we present Hierarchical Surrogate Rules (HSR), an algorithm that generates hierarchical rules based on user-defined parameters. We also contribute SuRE, a visual analytics (VA) system that integrates HSR and an interactive surrogate rule visualization, the Feature-Aligned Tree, which depicts rules as trees while aligning features for easier comparison. We evaluate the algorithm in terms of parameter sensitivity, time performance, and comparison with surrogate decision trees and find that it scales reasonably well and overcomes the shortcomings of surrogate decision trees. We evaluate the visualization and the system through a usability study and an observational study with domain experts. Our investigation shows that the participants can use feature-aligned trees to perform non-trivial tasks with very high accuracy. We also discuss many interesting findings, including a rule analysis task characterization, that can be used for visualization design and future research.

Analyzing classification model performance is a crucial task for machine learning practitioners. While practitioners often use count-based metrics derived from confusion matrices, like accuracy, many applications, such as weather prediction, sports betting, or patient risk prediction, rely on a classifier's predicted probabilities rather than predicted labels. In these instances, practitioners are concerned with producing a calibrated model, that is, one which outputs probabilities that reflect those of the true distribution. Model calibration is often analyzed visually, through static reliability diagrams, however, the traditional calibration visualization may suffer from a variety of drawbacks due to the strong aggregations it necessitates. Furthermore, count-based approaches are unable to sufficiently analyze model calibration. We present Calibrate, an interactive reliability diagram that addresses the aforementioned issues. Calibrate constructs a reliability diagram that is resistant to drawbacks in traditional approaches, and allows for interactive subgroup analysis and instance-level inspection. We demonstrate the utility of Calibrate through use cases on both real-world and synthetic data. We further validate Calibrate by presenting the results of a think-aloud experiment with data scientists who routinely analyze model calibration.

A deep learning algorithm is trained to predict wall-shear stress based on flow in the outer region of transitional and turbulent boundary layer flow. Flow variables sampled from the boundary layer region of eight high quality wall-resolved Large Eddy Simulations of transonic compressor cascades in the presence of shock-boundary layer interaction effects are used to train the model. About 1.1 TB of data stored in compressed text format generated from the LES calculations is used to perform the model training. The model is shown to be generic and able to predict complex boundary layer physics including laminar to turbulent transition caused by the shock-boundary layer interaction.