You’ve probably heard by now that machine learning algorithms can use big data to predict whether a donor will give to a charity, whether an infant in a NICU will develop sepsis, whether a customer will respond to an ad, and on and on. Machine learning can even drive cars and predict elections. ... Err, wait. Can it? I believe it can, but these recent high-profile hiccups should leave everyone who works with data (big or not) and machine learning algorithms asking themselves some very hard questions: do I understand my data? Do I understand the model and answers my machine learning algorithm is giving me? And do I trust these answers? Unfortunately, the complexity that bestows the extraordinary predictive abilities on machine learning algorithms also makes the answers the algorithms produce hard to understand, and maybe even hard to trust.
Although it is possible to enforce monotonicity constraints (a relationship that only changes in one direction) between independent variables and a machine-learned response function, machine learning algorithms tend to create nonlinear, non-monotonic, non-polynomial, and even non-continuous functions that approximate the relationship between independent and dependent variables in a data set. (This relationship might also be referred to as the conditional distribution of the dependent variables, given the values of the independent variables.) These functions can then make very specific predictions about the values of dependent variables for new data—whether a donor will give to a charity, an infant in a NICU will develop sepsis, a customer will respond to an ad, etc. Conversely, traditional linear models tend to create linear, monotonic, and continuous functions to approximate the very same relationships. Even though they’re not always the most accurate predictors, the elegant simplicity of linear models makes the results they generate easy to interpret.
While understanding and trusting models and results is a general requirement for good (data) science, model interpretability is a serious legal mandate in the regulated verticals of banking, insurance, and other industries. Business analysts, doctors, and industry researchers simply must understand and trust their models and modeling results. For this reason, linear models were the go-to applied predictive modeling tool for decades, even though it usually meant giving up a couple points on the accuracy scale. Today, many organizations and individuals are embracing machine learning algorithms for predictive modeling tasks, but difficulties in interpretation still present a barrier for the widespread, practical use of machine learning algorithms.
In this article, I present several approaches beyond the usual error measures and assessment plots for visualizing data and interpreting machine learning models and results. Users are encouraged to mix and match these techniques to best fit their own needs.
Wherever possible, “interpretability” of each technique in this article is deconstructed into more basic components—complexity, scope, understanding, and trust—which I will first outline below.
Complexity of response function to be explained
Linear, monotonic functions: Functions created by linear regression algorithms are probably the most interpretable class of models. These models will be referred to here as “linear and monotonic,” meaning that for a change in any given independent variable (or sometimes combination or function of an independent variable), the response function changes at a defined rate, in only one direction, and at a magnitude represented by a readily available coefficient. Monotonicity also enables intuitive and even automatic reasoning about predictions. For instance, if a lender rejects your credit card application, they can tell you why because their probability-of-default model often assumes your credit score, your account balances, and the length of your credit history are monotonically related to your ability to pay your credit card bill. When these explanations are created automatically, they are typically called “reason codes.” Of course, linear and monotonic response functions enable the calculation of relative variable importance measures, too. Linear and monotonic functions have several uses in machine learning interpretability. Part 1 and Part 2 below discuss the many ways linear, monotonic functions can be used to make machine learning interpretable.
Nonlinear, monotonic functions: Although most machine learned response functions are nonlinear, some can be constrained to be monotonic with respect to any given independent variable. While there is no single coefficient that represents the change in the response function induced by a change in a single independent variable, nonlinear and monotonic functions do always change in one direction as a single input variable changes. Nonlinear, monotonic response functions usually allow for the generation of both reason codes and relative variable importance measures. Nonlinear, monotonic response functions are highly interpretable and often suitable for use in regulated applications.
(Of course, there are linear, non-monotonic machine-learned response functions that can, for instance, be created by the multi-variate adaptive regression splines approach. These functions are not highlighted here because they tend to be less accurate predictors than purely nonlinear, non-monotonic functions, while also lacking the high interpretability of their monotonic counterparts.)
Nonlinear, non-monotonic functions: Most machine learning algorithms create nonlinear, non-monotonic response functions. This class of functions is the most difficult to interpret, as they can change in a positive and negative direction and at a varying rate for any change in an independent variable. Typically, the only standard interpretability measure these functions provide are relative variable importance measures. You may need to use a combination of additional techniques presented below to interpret these extremely complex models.
链接：
https://www.oreilly.com/ideas/ideas-on-interpreting-machine-learning
原文链接：
http://weibo.com/1402400261/EA5rj3JQV?from=page_1005051402400261_profile&wvr=6&mod=weibotime&type=comment#_rnd1489739687830
