Yin Lou

Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30-day Readmission

In machine learning often a tradeoff must be made between accuracy and intelligibility. More accurate models such as boosted trees, random forests, and neural nets usually are not intelligible, but more intelligible models such as logistic regression, naive-Bayes, and single decision trees often have significantly worse accuracy. This tradeoff sometimes limits the accuracy of models that can be applied in mission-critical applications such as healthcare where being able to understand, validate, edit, and trust a learned model is important. We present two case studies where high-performance generalized additive models with pairwise interactions (GA2Ms) are applied to real healthcare problems yielding intelligible models with state-of-the-art accuracy. In the pneumonia risk prediction case study, the intelligible model uncovers surprising patterns in the data that previously had prevented complex learned models from being fielded in this domain, but because it is intelligible and modular allows these patterns to be recognized and removed. In the 30-day hospital readmission case study, we show that the same methods scale to large datasets containing hundreds of thousands of patients and thousands of attributes while remaining intelligible and providing accuracy comparable to the best (unintelligible) machine learning methods.

Intelligible models for classification and regression

Complex models for regression and classification have high accuracy, but are unfortunately no longer interpretable by users. We study the performance of generalized additive models (GAMs), which combine single-feature models called shape functions through a linear function. Since the shape functions can be arbitrarily complex, GAMs are more accurate than simple linear models. But since they do not contain any interactions between features, they can be easily interpreted by users. We present the first large-scale empirical comparison of existing methods for learning GAMs. Our study includes existing spline and tree-based methods for shape functions and penalized least squares, gradient boosting, and backfitting for learning GAMs. We also present a new method based on tree ensembles with an adaptive number of leaves that consistently outperforms previous work. We complement our experimental results with a bias-variance analysis that explains how different shape models influence the additive model. Our experiments show that shallow bagged trees with gradient boosting distinguish itself as the best method on low- to medium-dimensional datasets.

Accurate intelligible models with pairwise interactions

Standard generalized additive models (GAMs) usually model the dependent variable as a sum of univariate models. Although previous studies have shown that standard GAMs can be interpreted by users, their accuracy is significantly less than more complex models that permit interactions. In this paper, we suggest adding selected terms of interacting pairs of features to standard GAMs. The resulting models, which we call GA2{M}

MatchMiner: efficient spanning structure mining in large image collections

-models, for Generalized Additive Models plus Interactions, consist of univariate terms and a small number of pairwise interaction terms. Since these models only include one- and two-dimensional components, the components of GA2M-models can be visualized and interpreted by users. To explore the huge (quadratic) number of pairs of features, we develop a novel, computationally efficient method called FAST for ranking all possible pairs of features as candidates for inclusion into the model. In a large-scale empirical study, we show the effectiveness of FAST in ranking candidate pairs of features. In addition, we show the surprising result that GA2M-models have almost the same performance as the best full-complexity models on a number of real datasets. Thus this paper postulates that for many problems, GA2M-models can yield models that are both intelligible and accurate.

ATLAS: a probabilistic algorithm for high dimensional similarity search

Many new computer vision applications are utilizing large-scale data- sets of places derived from the many billions of photos on the Web. Such applications often require knowledge of the visual connectivity structure of these image collections--describing which images overlap or are otherwise related--and an important step in understanding this structure is to identify connected components of this underlying image graph. As the structure of this graph is often initially unknown, this problem can be posed as one of exploring the connectivity between images as quickly as possible, by intelligently selecting a subset of image pairs for feature matching and geometric verification, without having to test all O(n2) possible pairs. We propose a novel, scalable algorithm called MatchMiner that efficiently explores visual relations between images, incorporating ideas from relevance feedback to improve decision making over time, as well as a simple yet effective rank distance measure for detecting outlier images. Using these ideas, our algorithm automatically prioritizes image pairs that can potentially connect or contribute to large connected components, using an information-theoretic algorithm to decide which image pairs to test next. Our experimental results show that MatchMiner can efficiently find connected components in large image collections, significantly outperforming state-of-the-art image matching methods.

Sparse Partially Linear Additive Models

Given a set of high dimensional binary vectors and a similarity function (such as Jaccard and Cosine), we study the problem of finding all pairs of vectors whose similarity exceeds a given threshold. The solution to this problem is a key component in many applications with feature-rich objects, such as text, images, music, videos, or social networks. In particular, there are many important emerging applications that require the use of relatively low similarity thresholds. We propose ATLAS, a probabilistic similarity search algorithm that in expectation finds a 1 - δ fraction of all similar vector pairs. ATLAS uses truly random permutations both to filter candidate pairs of vectors and to estimate the similarity between vectors. At a 97.5% recall rate, ATLAS consistently outperforms all state-of-the-art approaches and achieves a speed-up of up to two orders of magnitude over both exact and approximate algorithms.

Generating data series query workloads

The generalized partially linear additive model (GPLAM) is a flexible and interpretable approach to building predictive models. It combines features in an additive manner, allowing each to have either a linear or nonlinear effect on the response. However, the choice of which features to treat as linear or nonlinear is typically assumed known. Thus, to make a GPLAM a viable approach in situations in which little is known a priori about the features, one must overcome two primary model selection challenges: deciding which features to include in the model and determining which of these features to treat nonlinearly. We introduce the sparse partially linear additive model (SPLAM), which combines model fitting and both of these model selection challenges into a single convex optimization problem. SPLAM provides a bridge between the lasso and sparse additive models. Through a statistical oracle inequality and thorough simulation, we demonstrate that SPLAM can outperform other methods across a broad spectrum of statistical regimes, including the high-dimensional (p ≫ N) setting. We develop efficient algorithms that are applied to real datasets with half a million samples and over 45,000 features with excellent predictive performance. Supplementary materials for this article are available online.

Detecting Bias in Black-Box Models Using Transparent Model Distillation.

Data series (including time series) has attracted lots of interest in recent years. Most of the research has focused on how to efficiently support similarity or nearest neighbor queries over large data series collections (an important data mining task), and several data series summarization and indexing methods have been proposed in order to solve this problem. Up to this point, very little attention has been paid to properly evaluating such index structures, with most previous works relying solely on randomly selected data series to use as queries. In this work, we show that random workloads are inherently not suitable for the task at hand and we argue that there is a need for carefully generating query workloads. We define measures that capture the characteristics of queries, and we propose a method for generating workloads with the desired properties, that is, effectively evaluating and comparing data series summarizations and indexes. In our experimental evaluation, with carefully controlled query workloads, we shed light on key factors affecting the performance of nearest neighbor search in large data series collections. This is the first paper that introduces a method for quantifying hardness of data series queries, as well as the ability to generate queries of predefined hardness.