Su-In Lee

Hugh Chen, Ian C. Covert, Scott M. Lundberg et al.

Feature attributions based on the Shapley value are popular for explaining machine learning models; however, their estimation is complex from both a theoretical and computational standpoint. We disentangle this complexity into two factors: (1)~the approach to removing feature information, and (2)~the tractable estimation strategy. These two factors provide a natural lens through which we can better understand and compare 24 distinct algorithms. Based on the various feature removal approaches, we describe the multiple types of Shapley value feature attributions and methods to calculate each one. Then, based on the tractable estimation strategies, we characterize two distinct families of approaches: model-agnostic and model-specific approximations. For the model-agnostic approximations, we benchmark a wide class of estimation approaches and tie them to alternative yet equivalent characterizations of the Shapley value. For the model-specific approximations, we clarify the assumptions crucial to each method's tractability for linear, tree, and deep models. Finally, we identify gaps in the literature and promising future research directions.

Ian Covert, Wei Qiu, Mingyu Lu et al.

Feature selection helps reduce data acquisition costs in ML, but the standard approach is to train models with static feature subsets. Here, we consider the dynamic feature selection (DFS) problem where a model sequentially queries features based on the presently available information. DFS is often addressed with reinforcement learning, but we explore a simpler approach of greedily selecting features based on their conditional mutual information. This method is theoretically appealing but requires oracle access to the data distribution, so we develop a learning approach based on amortized optimization. The proposed method is shown to recover the greedy policy when trained to optimality, and it outperforms numerous existing feature selection methods in our experiments, thus validating it as a simple but powerful approach for this problem.

Ian Covert, Chanwoo Kim, Su-In Lee

Transformers have become a default architecture in computer vision, but understanding what drives their predictions remains a challenging problem. Current explanation approaches rely on attention values or input gradients, but these provide a limited view of a model's dependencies. Shapley values offer a theoretically sound alternative, but their computational cost makes them impractical for large, high-dimensional models. In this work, we aim to make Shapley values practical for vision transformers (ViTs). To do so, we first leverage an attention masking approach to evaluate ViTs with partial information, and we then develop a procedure to generate Shapley value explanations via a separate, learned explainer model. Our experiments compare Shapley values to many baseline methods (e.g., attention rollout, GradCAM, LRP), and we find that our approach provides more accurate explanations than existing methods for ViTs.

Ethan Weinberger, Ian Covert, Su-In Lee

Contrastive analysis (CA) refers to the exploration of variations uniquely enriched in a target dataset as compared to a corresponding background dataset generated from sources of variation that are irrelevant to a given task. For example, a biomedical data analyst may wish to find a small set of genes to use as a proxy for variations in genomic data only present among patients with a given disease (target) as opposed to healthy control subjects (background). However, as of yet the problem of feature selection in the CA setting has received little attention from the machine learning community. In this work we present contrastive feature selection (CFS), a method for performing feature selection in the CA setting. We motivate our approach with a novel information-theoretic analysis of representation learning in the CA setting, and we empirically validate CFS on a semi-synthetic dataset and four real-world biomedical datasets. We find that our method consistently outperforms previously proposed state-of-the-art supervised and fully unsupervised feature selection methods not designed for the CA setting. An open-source implementation of our method is available at https://github.com/suinleelab/CFS.

Mingyu Lu, Yushan Huang, Chris Lin et al.

As multi-agent systems (MAS) become increasingly complex, identifying the contributions of individual agents is critical for system optimization. However, existing approaches lack a rigorous, unified framework for credit assignment. In this work, we formalize agent attribution as a cooperative game, parameterized by the coalition distribution, removal protocol, and target metric. Using this framework, we show that Leave-One-Out (LOO) identifies bottleneck agents as effectively as combinatorial methods, but at a fraction of the computational cost. We also demonstrate that removal protocols induce distinct games: Agent ablation isolates structural bottlenecks, whereas introspective LLM judges fail to faithfully approximate this behavior. Furthermore, to evaluate the utility of specific agent backbones, we introduce attribution via model replacement. By substituting underlying models of low-contribution agents, we improve task performance by up to 17% while reducing cost by up to 35% across three benchmarks. Finally, we apply our framework to audit a medical MAS, revealing that agent contributions to diagnostic accuracy and ethical behavior are often decoupled. By intervening on counterproductive roles, we observe an increase in ethics alignment while maintaining diagnostic accuracy. Overall, this work provides a principled approach for cost-effective MAS attribution and intervention.

Chris Lin, Ian Covert, Su-In Lee

To explain predictions made by complex machine learning models, many feature attribution methods have been developed that assign importance scores to input features. Some recent work challenges the robustness of these methods by showing that they are sensitive to input and model perturbations, while other work addresses this issue by proposing robust attribution methods. However, previous work on attribution robustness has focused primarily on gradient-based feature attributions, whereas the robustness of removal-based attribution methods is not currently well understood. To bridge this gap, we theoretically characterize the robustness properties of removal-based feature attributions. Specifically, we provide a unified analysis of such methods and derive upper bounds for the difference between intact and perturbed attributions, under settings of both input and model perturbations. Our empirical results on synthetic and real-world data validate our theoretical results and demonstrate their practical implications, including the ability to increase attribution robustness by improving the model's Lipschitz regularity.

Soham Gadgil, Ian Covert, Su-In Lee

Dynamic feature selection, where we sequentially query features to make accurate predictions with a minimal budget, is a promising paradigm to reduce feature acquisition costs and provide transparency into a model's predictions. The problem is challenging, however, as it requires both predicting with arbitrary feature sets and learning a policy to identify valuable selections. Here, we take an information-theoretic perspective and prioritize features based on their mutual information with the response variable. The main challenge is implementing this policy, and we design a new approach that estimates the mutual information in a discriminative rather than generative fashion. Building on our approach, we then introduce several further improvements: allowing variable feature budgets across samples, enabling non-uniform feature costs, incorporating prior information, and exploring modern architectures to handle partial inputs. Our experiments show that our method provides consistent gains over recent methods across a variety of datasets.

Chris Lin, Hugh Chen, Chanwoo Kim et al.

Despite the widespread use of unsupervised models, very few methods are designed to explain them. Most explanation methods explain a scalar model output. However, unsupervised models output representation vectors, the elements of which are not good candidates to explain because they lack semantic meaning. To bridge this gap, recent works defined a scalar explanation output: a dot product-based similarity in the representation space to the sample being explained (i.e., an explicand). Although this enabled explanations of unsupervised models, the interpretation of this approach can still be opaque because similarity to the explicand's representation may not be meaningful to humans. To address this, we propose contrastive corpus similarity, a novel and semantically meaningful scalar explanation output based on a reference corpus and a contrasting foil set of samples. We demonstrate that contrastive corpus similarity is compatible with many post-hoc feature attribution methods to generate COntrastive COrpus Attributions (COCOA) and quantitatively verify that features important to the corpus are identified. We showcase the utility of COCOA in two ways: (i) we draw insights by explaining augmentations of the same image in a contrastive learning setting (SimCLR); and (ii) we perform zero-shot object localization by explaining the similarity of image representations to jointly learned text representations (CLIP).

Soham Gadgil, Chris Lin, Su-In Lee

Steering vectors have emerged as a lightweight and effective approach for aligning large language models (LLMs) at inference time, enabling modulation over model behaviors by shifting LLM representations towards a target behavior. However, existing methods typically apply steering vectors at a globally fixed layer, implicitly assuming that the optimal intervention layer is invariant across inputs. We argue that this assumption is fundamentally limited, as representations relevant to a target behavior can be encoded at different layers depending on the input. Theoretically, we show that different inputs can require steering at different layers to achieve alignment with a desirable model behavior. We also provide empirical evidence that the optimal steering layer varies substantially across inputs in practice. Motivated by these observations, we introduce Where to Steer (W2S), a framework that adaptively selects the intervention layer conditioned on the input, by learning a mapping from input embeddings to optimal steering layers. Across multiple LLMs and alignment behaviors, W2S consistently outperforms fixed-layer baselines, with improvements in both in-distribution and out-of-distribution settings. Our findings highlight the importance of input-dependent control in LLM alignment and demonstrate that adaptive layer selection is a key design dimension missing in the current methodology of steering vectors.

Joseph D. Janizek, Pascal Sturmfels, Su-In Lee

Recent work has shown great promise in explaining neural network behavior. In particular, feature attribution methods explain which features were most important to a model's prediction on a given input. However, for many tasks, simply knowing which features were important to a model's prediction may not provide enough insight to understand model behavior. The interactions between features within the model may better help us understand not only the model, but also why certain features are more important than others. In this work, we present Integrated Hessians, an extension of Integrated Gradients that explains pairwise feature interactions in neural networks. Integrated Hessians overcomes several theoretical limitations of previous methods to explain interactions, and unlike such previous methods is not limited to a specific architecture or class of neural network. Additionally, we find that our method is faster than existing methods when the number of features is large, and outperforms previous methods on existing quantitative benchmarks. Code available at https://github.com/suinleelab/path_explain

Joseph D. Janizek, Gabriel Erion, Alex J. DeGrave et al.

While deep learning has shown promise in the domain of disease classification from medical images, models based on state-of-the-art convolutional neural network architectures often exhibit performance loss due to dataset shift. Models trained using data from one hospital system achieve high predictive performance when tested on data from the same hospital, but perform significantly worse when they are tested in different hospital systems. Furthermore, even within a given hospital system, deep learning models have been shown to depend on hospital- and patient-level confounders rather than meaningful pathology to make classifications. In order for these models to be safely deployed, we would like to ensure that they do not use confounding variables to make their classification, and that they will work well even when tested on images from hospitals that were not included in the training data. We attempt to address this problem in the context of pneumonia classification from chest radiographs. We propose an approach based on adversarial optimization, which allows us to learn more robust models that do not depend on confounders. Specifically, we demonstrate improved out-of-hospital generalization performance of a pneumonia classifier by training a model that is invariant to the view position of chest radiographs (anterior-posterior vs. posterior-anterior). Our approach leads to better predictive performance on external hospital data than both a standard baseline and previously proposed methods to handle confounding, and also suggests a method for identifying models that may rely on confounders. Code available at https://github.com/suinleelab/cxr_adv.

Scott M. Lundberg, Gabriel G. Erion, Su-In Lee

Interpreting predictions from tree ensemble methods such as gradient boosting machines and random forests is important, yet feature attribution for trees is often heuristic and not individualized for each prediction. Here we show that popular feature attribution methods are inconsistent, meaning they can lower a feature's assigned importance when the true impact of that feature actually increases. This is a fundamental problem that casts doubt on any comparison between features. To address it we turn to recent applications of game theory and develop fast exact tree solutions for SHAP (SHapley Additive exPlanation) values, which are the unique consistent and locally accurate attribution values. We then extend SHAP values to interaction effects and define SHAP interaction values. We propose a rich visualization of individualized feature attributions that improves over classic attribution summaries and partial dependence plots, and a unique "supervised" clustering (clustering based on feature attributions). We demonstrate better agreement with human intuition through a user study, exponential improvements in run time, improved clustering performance, and better identification of influential features. An implementation of our algorithm has also been merged into XGBoost and LightGBM, see http://github.com/slundberg/shap for details.

Mingyu Lu, Soham Gadgil, Chris Lin et al.

As Text-to-Image (T2I) diffusion models are increasingly used in real-world creative workflows, a principled framework for valuing contributors who provide a collection of data is essential for fair compensation and sustainable data marketplaces. While the Shapley value offers a theoretically grounded approach to attribution, it faces a dual computational bottleneck: (i) the prohibitive cost of exhaustive model retraining for each sampled subset of players (i.e., data contributors) and (ii) the combinatorial number of subsets needed to estimate marginal contributions due to contributor interactions. To this end, we propose SurrogateSHAP, a retraining-free framework that approximates the expensive retraining game through inference from a pretrained model. To further improve efficiency, we employ a gradient-boosted tree to approximate the utility function and derive Shapley values analytically from the tree-based model. We evaluate SurrogateSHAP across three diverse attribution tasks: (i) image quality for DDPM-CFG on CIFAR-20, (ii) aesthetics for Stable Diffusion on Post-Impressionist artworks, and (iii) product diversity for FLUX.1 on Fashion-Product data. Across settings, SurrogateSHAP outperforms prior methods while substantially reducing computational overhead, consistently identifying influential contributors across multiple utility metrics. Finally, we demonstrate that SurrogateSHAP effectively localizes data sources responsible for spurious correlations in clinical images, providing a scalable path toward auditing safety-critical generative models.

Mingyu Lu, Yifang Chen, Su-In Lee

Learning personalized cancer treatment with machine learning holds great promise to improve cancer patients' chance of survival. Despite recent advances in machine learning and precision oncology, this approach remains challenging as collecting data in preclinical/clinical studies for modeling multiple treatment efficacies is often an expensive, time-consuming process. Moreover, the randomization in treatment allocation proves to be suboptimal since some participants/samples are not receiving the most appropriate treatments during the trial. To address this challenge, we formulate drug screening study as a "contextual bandit" problem, in which an algorithm selects anticancer therapeutics based on contextual information about cancer cell lines while adapting its treatment strategy to maximize treatment response in an "online" fashion. We propose using a novel deep Bayesian bandits framework that uses functional prior to approximate posterior for drug response prediction based on multi-modal information consisting of genomic features and drug structure. We empirically evaluate our method on three large-scale in vitro pharmacogenomic datasets and show that our approach outperforms several benchmarks in identifying optimal treatment for a given cell line.

Ian Covert, Chanwoo Kim, Su-In Lee et al.

Many tasks in explainable machine learning, such as data valuation and feature attribution, perform expensive computation for each data point and are intractable for large datasets. These methods require efficient approximations, and although amortizing the process by learning a network to directly predict the desired output is a promising solution, training such models with exact labels is often infeasible. We therefore explore training amortized models with noisy labels, and we find that this is inexpensive and surprisingly effective. Through theoretical analysis of the label noise and experiments with various models and datasets, we show that this approach tolerates high noise levels and significantly accelerates several feature attribution and data valuation methods, often yielding an order of magnitude speedup over existing approaches.

Soham Gadgil, Chris Lin, Su-In Lee

Sparse autoencoders (SAEs) are used to decompose neural network activations into human-interpretable features. Typically, features learned by a single SAE are used for downstream applications. However, it has recently been shown that SAEs trained with different initial weights can learn different features, demonstrating that a single SAE captures only a limited subset of features that can be extracted from the activation space. Motivated by this limitation, we propose to ensemble multiple SAEs through naive bagging and boosting. Specifically, SAEs trained with different weight initializations are ensembled in naive bagging, whereas SAEs sequentially trained to minimize the residual error are ensembled in boosting. We evaluate our ensemble approaches with three settings of language models and SAE architectures. Our empirical results demonstrate that ensembling SAEs can improve the reconstruction of language model activations, diversity of features, and SAE stability. Furthermore, ensembling SAEs performs better than applying a single SAE on downstream tasks such as concept detection and spurious correlation removal, showing improved practical utility.

Chris Lin, Mingyu Lu, Chanwoo Kim et al.

As diffusion models are deployed in real-world settings, and their performance is driven by training data, appraising the contribution of data contributors is crucial to creating incentives for sharing quality data and to implementing policies for data compensation. Depending on the use case, model performance corresponds to various global properties of the distribution learned by a diffusion model (e.g., overall aesthetic quality). Hence, here we address the problem of attributing global properties of diffusion models to data contributors. The Shapley value provides a principled approach to valuation by uniquely satisfying game-theoretic axioms of fairness. However, estimating Shapley values for diffusion models is computationally impractical because it requires retraining on many training data subsets corresponding to different contributors and rerunning inference. We introduce a method to efficiently retrain and rerun inference for Shapley value estimation, by leveraging model pruning and fine-tuning. We evaluate the utility of our method with three use cases: (i) image quality for a DDPM trained on a CIFAR dataset, (ii) demographic diversity for an LDM trained on CelebA-HQ, and (iii) aesthetic quality for a Stable Diffusion model LoRA-finetuned on Post-Impressionist artworks. Our results empirically demonstrate that our framework can identify important data contributors across models' global properties, outperforming existing attribution methods for diffusion models.

Ethan Weinberger, Nicasia Beebe-Wang, Su-In Lee

In the contrastive analysis (CA) setting, machine learning practitioners are specifically interested in discovering patterns that are enriched in a target dataset as compared to a background dataset generated from sources of variation irrelevant to the task at hand. For example, a biomedical data analyst may seek to understand variations in genomic data only present among patients with a given disease as opposed to those also present in healthy control subjects. Such scenarios have motivated the development of contrastive latent variable models to isolate variations unique to these target datasets from those shared across the target and background datasets, with current state of the art models based on the variational autoencoder (VAE) framework. However, previously proposed models do not explicitly enforce the constraints on latent variables underlying CA, potentially leading to the undesirable leakage of information between the two sets of latent variables. Here we propose the moment matching contrastive VAE (MM-cVAE), a reformulation of the VAE for CA that uses the maximum mean discrepancy to explicitly enforce two crucial latent variable constraints underlying CA. On three challenging CA tasks we find that our method outperforms the previous state-of-the-art both qualitatively and on a set of quantitative metrics.

Neil Jethani, Mukund Sudarshan, Ian Covert et al.

Shapley values are widely used to explain black-box models, but they are costly to calculate because they require many model evaluations. We introduce FastSHAP, a method for estimating Shapley values in a single forward pass using a learned explainer model. FastSHAP amortizes the cost of explaining many inputs via a learning approach inspired by the Shapley value's weighted least squares characterization, and it can be trained using standard stochastic gradient optimization. We compare FastSHAP to existing estimation approaches, revealing that it generates high-quality explanations with orders of magnitude speedup.

Sahil Verma, Aditya Lahiri, John P. Dickerson et al.

As machine learning (ML) systems take a more prominent and central role in contributing to life-impacting decisions, ensuring their trustworthiness and accountability is of utmost importance. Explanations sit at the core of these desirable attributes of a ML system. The emerging field is frequently called ``Explainable AI (XAI)'' or ``Explainable ML.'' The goal of explainable ML is to intuitively explain the predictions of a ML system, while adhering to the needs to various stakeholders. Many explanation techniques were developed with contributions from both academia and industry. However, there are several existing challenges that have not garnered enough interest and serve as roadblocks to widespread adoption of explainable ML. In this short paper, we enumerate challenges in explainable ML from an industry perspective. We hope these challenges will serve as promising future research directions, and would contribute to democratizing explainable ML.

Hugh Chen, Scott M. Lundberg, Su-In Lee

Local feature attribution methods are increasingly used to explain complex machine learning models. However, current methods are limited because they are extremely expensive to compute or are not capable of explaining a distributed series of models where each model is owned by a separate institution. The latter is particularly important because it often arises in finance where explanations are mandated. Here, we present DeepSHAP, a tractable method to propagate local feature attributions through complex series of models based on a connection to the Shapley value. We evaluate DeepSHAP across biological, health, and financial datasets to show that it provides equally salient explanations an order of magnitude faster than existing model-agnostic attribution techniques and demonstrate its use in an important distributed series of models setting.

Ian Covert, Su-In Lee

The Shapley value concept from cooperative game theory has become a popular technique for interpreting ML models, but efficiently estimating these values remains challenging, particularly in the model-agnostic setting. Here, we revisit the idea of estimating Shapley values via linear regression to understand and improve upon this approach. By analyzing the original KernelSHAP alongside a newly proposed unbiased version, we develop techniques to detect its convergence and calculate uncertainty estimates. We also find that the original version incurs a negligible increase in bias in exchange for significantly lower variance, and we propose a variance reduction technique that further accelerates the convergence of both estimators. Finally, we develop a version of KernelSHAP for stochastic cooperative games that yields fast new estimators for two global explanation methods.

Ian Covert, Scott Lundberg, Su-In Lee

Researchers have proposed a wide variety of model explanation approaches, but it remains unclear how most methods are related or when one method is preferable to another. We describe a new unified class of methods, removal-based explanations, that are based on the principle of simulating feature removal to quantify each feature's influence. These methods vary in several respects, so we develop a framework that characterizes each method along three dimensions: 1) how the method removes features, 2) what model behavior the method explains, and 3) how the method summarizes each feature's influence. Our framework unifies 26 existing methods, including several of the most widely used approaches: SHAP, LIME, Meaningful Perturbations, and permutation tests. This newly understood class of explanation methods has rich connections that we examine using tools that have been largely overlooked by the explainability literature. To anchor removal-based explanations in cognitive psychology, we show that feature removal is a simple application of subtractive counterfactual reasoning. Ideas from cooperative game theory shed light on the relationships and trade-offs among different methods, and we derive conditions under which all removal-based explanations have information-theoretic interpretations. Through this analysis, we develop a unified framework that helps practitioners better understand model explanation tools, and that offers a strong theoretical foundation upon which future explainability research can build.

Ian Covert, Scott Lundberg, Su-In Lee

Researchers have proposed a wide variety of model explanation approaches, but it remains unclear how most methods are related or when one method is preferable to another. We examine the literature and find that many methods are based on a shared principle of explaining by removing - essentially, measuring the impact of removing sets of features from a model. These methods vary in several respects, so we develop a framework for removal-based explanations that characterizes each method along three dimensions: 1) how the method removes features, 2) what model behavior the method explains, and 3) how the method summarizes each feature's influence. Our framework unifies 26 existing methods, including several of the most widely used approaches (SHAP, LIME, Meaningful Perturbations, permutation tests). Exposing the fundamental similarities between these methods empowers users to reason about which tools to use, and suggests promising directions for ongoing model explainability research.

Hugh Chen, Joseph D. Janizek, Scott Lundberg et al.

A variety of recent papers discuss the application of Shapley values, a concept for explaining coalitional games, for feature attribution in machine learning. However, the correct way to connect a machine learning model to a coalitional game has been a source of controversy. The two main approaches that have been proposed differ in the way that they condition on known features, using either (1) an interventional or (2) an observational conditional expectation. While previous work has argued that one of the two approaches is preferable in general, we argue that the choice is application dependent. Furthermore, we argue that the choice comes down to whether it is desirable to be true to the model or true to the data. We use linear models to investigate this choice. After deriving an efficient method for calculating observational conditional expectation Shapley values for linear models, we investigate how correlation in simulated data impacts the convergence of observational conditional expectation Shapley values. Finally, we present two real data examples that we consider to be representative of possible use cases for feature attribution -- (1) credit risk modeling and (2) biological discovery. We show how a different choice of value function performs better in each scenario, and how possible attributions are impacted by modeling choices.

Ian Covert, Scott Lundberg, Su-In Lee

Understanding the inner workings of complex machine learning models is a long-standing problem and most recent research has focused on local interpretability. To assess the role of individual input features in a global sense, we explore the perspective of defining feature importance through the predictive power associated with each feature. We introduce two notions of predictive power (model-based and universal) and formalize this approach with a framework of additive importance measures, which unifies numerous methods in the literature. We then propose SAGE, a model-agnostic method that quantifies predictive power while accounting for feature interactions. Our experiments show that SAGE can be calculated efficiently and that it assigns more accurate importance values than other methods.

Hugh Chen, Scott Lundberg, Gabe Erion et al.

Hundreds of millions of surgical procedures take place annually across the world, which generate a prevalent type of electronic health record (EHR) data comprising time series physiological signals. Here, we present a transferable embedding method (i.e., a method to transform time series signals into input features for predictive machine learning models) named PHASE (PHysiologicAl Signal Embeddings) that enables us to more accurately forecast adverse surgical outcomes based on physiological signals. We evaluate PHASE on minute-by-minute EHR data of more than 50,000 surgeries from two operating room (OR) datasets and patient stays in an intensive care unit (ICU) dataset. PHASE outperforms other state-of-the-art approaches, such as long-short term memory networks trained on raw data and gradient boosted trees trained on handcrafted features, in predicting five distinct outcomes: hypoxemia, hypocapnia, hypotension, hypertension, and phenylephrine administration. In a transfer learning setting where we train embedding models in one dataset then embed signals and predict adverse events in unseen data, PHASE achieves significantly higher prediction accuracy at lower computational cost compared to conventional approaches. Finally, given the importance of understanding models in clinical applications we demonstrate that PHASE is explainable and validate our predictive models using local feature attribution methods.

Ethan Weinberger, Joseph Janizek, Su-In Lee

Feature attribution methods, which explain an individual prediction made by a model as a sum of attributions for each input feature, are an essential tool for understanding the behavior of complex deep learning models. However, ensuring that models produce meaningful explanations, rather than ones that rely on noise, is not straightforward. Exacerbating this problem is the fact that attribution methods do not provide insight as to why features are assigned their attribution values, leading to explanations that are difficult to interpret. In real-world problems we often have sets of additional information for each feature that are predictive of that feature's importance to the task at hand. Here, we propose the deep attribution prior (DAPr) framework to exploit such information to overcome the limitations of attribution methods. Our framework jointly learns a relationship between prior information and feature importance, as well as biases models to have explanations that rely on features predicted to be important. We find that our framework both results in networks that generalize better to out of sample data and admits new methods for interpreting model behavior.

Hugh Chen, Scott Lundberg, Su-In Lee

In healthcare, making the best possible predictions with complex models (e.g., neural networks, ensembles/stacks of different models) can impact patient welfare. In order to make these complex models explainable, we present DeepSHAP for mixed model types, a framework for layer wise propagation of Shapley values that builds upon DeepLIFT (an existing approach for explaining neural networks). We show that in addition to being able to explain neural networks, this new framework naturally enables attributions for stacks of mixed models (e.g., neural network feature extractor into a tree model) as well as attributions of the loss. Finally, we theoretically justify a method for obtaining attributions with respect to a background distribution (under a Shapley value framework).

Gabriel Erion, Joseph D. Janizek, Pascal Sturmfels et al.

Recent research has demonstrated that feature attribution methods for deep networks can themselves be incorporated into training; these attribution priors optimize for a model whose attributions have certain desirable properties -- most frequently, that particular features are important or unimportant. These attribution priors are often based on attribution methods that are not guaranteed to satisfy desirable interpretability axioms, such as completeness and implementation invariance. Here, we introduce attribution priors to optimize for higher-level properties of explanations, such as smoothness and sparsity, enabled by a fast new attribution method formulation called expected gradients that satisfies many important interpretability axioms. This improves model performance on many real-world tasks where previous attribution priors fail. Our experiments show that the gains from combining higher-level attribution priors with expected gradients attributions are consistent across image, gene expression, and health care data sets. We believe this work motivates and provides the necessary tools to support the widespread adoption of axiomatic attribution priors in many areas of applied machine learning. The implementations and our results have been made freely available to academic communities.

Scott M. Lundberg, Gabriel Erion, Hugh Chen et al.

Tree-based machine learning models such as random forests, decision trees, and gradient boosted trees are the most popular non-linear predictive models used in practice today, yet comparatively little attention has been paid to explaining their predictions. Here we significantly improve the interpretability of tree-based models through three main contributions: 1) The first polynomial time algorithm to compute optimal explanations based on game theory. 2) A new type of explanation that directly measures local feature interaction effects. 3) A new set of tools for understanding global model structure based on combining many local explanations of each prediction. We apply these tools to three medical machine learning problems and show how combining many high-quality local explanations allows us to represent global structure while retaining local faithfulness to the original model. These tools enable us to i) identify high magnitude but low frequency non-linear mortality risk factors in the general US population, ii) highlight distinct population sub-groups with shared risk characteristics, iii) identify non-linear interaction effects among risk factors for chronic kidney disease, and iv) monitor a machine learning model deployed in a hospital by identifying which features are degrading the model's performance over time. Given the popularity of tree-based machine learning models, these improvements to their interpretability have implications across a broad set of domains.

Hugh Chen, Scott Lundberg, Su-In Lee

Time series data constitutes a distinct and growing problem in machine learning. As the corpus of time series data grows larger, deep models that simultaneously learn features and classify with these features can be intractable or suboptimal. In this paper, we present feature learning via long short term memory (LSTM) networks and prediction via gradient boosting trees (XGB). Focusing on the consequential setting of electronic health record data, we predict the occurrence of hypoxemia five minutes into the future based on past features. We make two observations: 1) long short term memory networks are effective at capturing long term dependencies based on a single feature and 2) gradient boosting trees are capable of tractably combining a large number of features including static features like height and weight. With these observations in mind, we generate features by performing "supervised" representation learning with LSTM networks. Augmenting the original XGB model with these features gives significantly better performance than either individual method.

Gabriel Erion, Hugh Chen, Scott M. Lundberg et al.

We use a deep learning model trained only on a patient's blood oxygenation data (measurable with an inexpensive fingertip sensor) to predict impending hypoxemia (low blood oxygen) more accurately than trained anesthesiologists with access to all the data recorded in a modern operating room. We also provide a simple way to visualize the reason why a patient's risk is low or high by assigning weight to the patient's past blood oxygen values. This work has the potential to provide cutting-edge clinical decision support in low-resource settings, where rates of surgical complication and death are substantially greater than in high-resource areas.

Hugh Chen, Scott Lundberg, Su-In Lee

We present the checkpoint ensembles method that can learn ensemble models on a single training process. Although checkpoint ensembles can be applied to any parametric iterative learning technique, here we focus on neural networks. Neural networks' composable and simple neurons make it possible to capture many individual and interaction effects among features. However, small sample sizes and sampling noise may result in patterns in the training data that are not representative of the true relationship between the features and the outcome. As a solution, regularization during training is often used (e.g. dropout). However, regularization is no panacea -- it does not perfectly address overfitting. Even with methods like dropout, two methodologies are commonly used in practice. First is to utilize a validation set independent to the training set as a way to decide when to stop training. Second is to use ensemble methods to further reduce overfitting and take advantage of local optima (i.e. averaging over the predictions of several models). In this paper, we explore checkpoint ensembles -- a simple technique that combines these two ideas in one training process. Checkpoint ensembles improve performance by averaging the predictions from "checkpoints" of the best models within single training process. We use three real-world data sets -- text, image, and electronic health record data -- using three prediction models: a vanilla neural network, a convolutional neural network, and a long short term memory network to show that checkpoint ensembles outperform existing methods: a method that selects a model by minimum validation score, and two methods that average models by weights. Our results also show that checkpoint ensembles capture a portion of the performance gains that traditional ensembles provide.

Scott M. Lundberg, Su-In Lee

Note that a newer expanded version of this paper is now available at: arXiv:1802.03888 It is critical in many applications to understand what features are important for a model, and why individual predictions were made. For tree ensemble methods these questions are usually answered by attributing importance values to input features, either globally or for a single prediction. Here we show that current feature attribution methods are inconsistent, which means changing the model to rely more on a given feature can actually decrease the importance assigned to that feature. To address this problem we develop fast exact solutions for SHAP (SHapley Additive exPlanation) values, which were recently shown to be the unique additive feature attribution method based on conditional expectations that is both consistent and locally accurate. We integrate these improvements into the latest version of XGBoost, demonstrate the inconsistencies of current methods, and show how using SHAP values results in significantly improved supervised clustering performance. Feature importance values are a key part of understanding widely used models such as gradient boosting trees and random forests, so improvements to them have broad practical implications.

Scott Lundberg, Su-In Lee

Understanding why a model makes a certain prediction can be as crucial as the prediction's accuracy in many applications. However, the highest accuracy for large modern datasets is often achieved by complex models that even experts struggle to interpret, such as ensemble or deep learning models, creating a tension between accuracy and interpretability. In response, various methods have recently been proposed to help users interpret the predictions of complex models, but it is often unclear how these methods are related and when one method is preferable over another. To address this problem, we present a unified framework for interpreting predictions, SHAP (SHapley Additive exPlanations). SHAP assigns each feature an importance value for a particular prediction. Its novel components include: (1) the identification of a new class of additive feature importance measures, and (2) theoretical results showing there is a unique solution in this class with a set of desirable properties. The new class unifies six existing methods, notable because several recent methods in the class lack the proposed desirable properties. Based on insights from this unification, we present new methods that show improved computational performance and/or better consistency with human intuition than previous approaches.

Scott Lundberg, Su-In Lee

Understanding why a model made a certain prediction is crucial in many data science fields. Interpretable predictions engender appropriate trust and provide insight into how the model may be improved. However, with large modern datasets the best accuracy is often achieved by complex models even experts struggle to interpret, which creates a tension between accuracy and interpretability. Recently, several methods have been proposed for interpreting predictions from complex models by estimating the importance of input features. Here, we present how a model-agnostic additive representation of the importance of input features unifies current methods. This representation is optimal, in the sense that it is the only set of additive values that satisfies important properties. We show how we can leverage these properties to create novel visual explanations of model predictions. The thread of unity that this representation weaves through the literature indicates that there are common principles to be learned about the interpretation of model predictions that apply in many scenarios.

Kean Ming Tan, Palma London, Karthik Mohan et al.

We consider the problem of learning a high-dimensional graphical model in which certain hub nodes are highly-connected to many other nodes. Many authors have studied the use of an l1 penalty in order to learn a sparse graph in high-dimensional setting. However, the l1 penalty implicitly assumes that each edge is equally likely and independent of all other edges. We propose a general framework to accommodate more realistic networks with hub nodes, using a convex formulation that involves a row-column overlap norm penalty. We apply this general framework to three widely-used probabilistic graphical models: the Gaussian graphical model, the covariance graph model, and the binary Ising model. An alternating direction method of multipliers algorithm is used to solve the corresponding convex optimization problems. On synthetic data, we demonstrate that our proposed framework outperforms competitors that do not explicitly model hub nodes. We illustrate our proposal on a webpage data set and a gene expression data set.

Karthik Mohan, Palma London, Maryam Fazel et al.

We consider the problem of estimating high-dimensional Gaussian graphical models corresponding to a single set of variables under several distinct conditions. This problem is motivated by the task of recovering transcriptional regulatory networks on the basis of gene expression data {containing heterogeneous samples, such as different disease states, multiple species, or different developmental stages}. We assume that most aspects of the conditional dependence networks are shared, but that there are some structured differences between them. Rather than assuming that similarities and differences between networks are driven by individual edges, we take a node-based approach, which in many cases provides a more intuitive interpretation of the network differences. We consider estimation under two distinct assumptions: (1) differences between the K networks are due to individual nodes that are perturbed across conditions, or (2) similarities among the K networks are due to the presence of common hub nodes that are shared across all K networks. Using a row-column overlap norm penalty function, we formulate two convex optimization problems that correspond to these two assumptions. We solve these problems using an alternating direction method of multipliers algorithm, and we derive a set of necessary and sufficient conditions that allows us to decompose the problem into independent subproblems so that our algorithm can be scaled to high-dimensional settings. Our proposal is illustrated on synthetic data, a webpage data set, and a brain cancer gene expression data set.