Variable Importance Assessment in Random Forest Regressions

[ machine-learning  random-forest  easi  ]

Today, I’m reading through 2009’s Variable Importance Assessment in Regression: Linear Regression versus Random Forest (at the time of this writing, academia.edu had a pdf of this paper)

In what follows are quotes and notes.

On Variable Importance (Philosophy)

“Variable importance in regression is an important topic in applied statistics that keeps coming up in spite of critics who basically claim that the question should not have been asked in the first place.”

“Variable importance is not very well defined as a concept. Even in the well-developed linear model and the standard n»p situation, there is no theoretically defined variable importance metric in the sense of a parametric quantity that a variable importance estimator should try to estimate.”

“In the absence of a clearly agreed true value, ad hoc proposals for empirical assessment of variable importance have been made, and desirability criteria for these have been formulated, for example, ‘decomposition’ of R2 into ‘nonnegative contributions attributable to each regressor’ has been postulated. Popular approaches for empirically assessing a variable’s importance include squared correlations (a completely marginal approach) and squared standardized coefficients (an approach conditional on all other variables in the model).”

Given no clear definition of variable importance, and the plethora of choices available to define such a thing, one must be prudent. “Depending on the research question at hand, the focus of the variable importance assessment in regression can be explanatory or predictive importance or a mixture of both.” What definition best suits your goals? (Heck, you can use more than one, as long as you clearly know what each measure means and what are its pros and cons.)

For an intuition concerning what constitutes an explanatory measure of importance versus a predictive measure of importance, the authors provide two chains:

  1. X2 -> X1 -> Y
  2. X2 <- X1 -> Y

Assuming all relationships are linear, the authors discuss that two variable importances often used with linear models (e.g., squares of standardized coefficients) would zero out the importance of X2 for both causal chains since, conditional on X1, X2 does not contribute any additional information to the prediction. Given this, do you consider this type of importance explanatory or predictive? It is neither, at least not purely. It makes sense for predictive purposes, but doesn’t tell the whole story (e.g., if one removes X1, then a prediction is still possible as X2 will take X1’s place). For explanatory purposes, it makes sense for causal chain (ii), since we see in that chain that X2 is only correlated with Y since both are driven by X1; thus, X2 should not be considered an important explanatory/causal variable for Y. However, for causal chain (i), this measure of variable importance is not explanatory: X2 very clearly drives Y, thus should not be considered to have zero explanatory importance! In practice, a linear model does not actually differentiate which of the two causal chains are at play. So the variable importance might be accidentally explanatory in the second causal chain, but not generally explanatory. It more generally a measure of predictive power, though as mentioned, just one story of predictive power among many.

Linear Regression vs Random Forest

“Linear regression is a classical parametric method which requires explicit modeling of nonlinearities and interactions, if necessary. It is known to be reasonably robust, if the number of observations n is distinctly larger than the number of variables p (n>p). With more variables than observations (p>n or even p»n), linear regression breaks down, unless shrinkage methods are used like ridge regression, the lasso, or the elastic net as a combination of both.”

“Random forests, on the other hand, are nonparametric and allow nonlinearities and interactions to be learned from the data without any need to explicitly model them. Also, they have been reported to work well not only for the n»p setting but also for data mining inthe p»n setting.”

What is a Random Forest anyway?

“A random forest is random in two ways: (i) each tree is based on a random subset of the observations, and (ii) each split within each tree is created based on a random subset of mtry candidate variables. Trees are quite unstable, so that this randomness creates differences in individual trees’ predictions. The overall prediction of the forest is the average of predictions from the individual trees—because individual trees produce multidimensional step functions, their average is again a multidimensional step function that can nevertheless predict smooth functions because it aggregates a large number of different trees.”

Improving Upon the Original Random Forest

“The splitting approach in CART trees has been known for a long time to be unfair in the presence of regressor variables of different types, categorical variables with different numbers of categories, or differing numbers of missing values (cf., e.g.,Breiman 1984; Shih and Tsai 2004). To avoid this variable selection bias, Hothorn, Hornik, and Zeileis (2006b) proposed to use multiplicity-adjusted conditional tests rather than maximum impurity reduction as the splitting criterion.”

  • Wow, this quote would have been helpful a few months ago :-p
  • Also, note that I don’t think their (interesting) p-value procedure at each node is implemented in sklearn (if anywhere (?))
  • In most places (sklearn included), a variation on what this paper refers to as RF-CART is implemented (true RF-CART grows individual trees to node purity, whereas usually an API allows you to make a few choices about how the trees are grown). The paper refers to the p-value-split trees as conditional inference (CI) trees, and random forests built up of CI-trees as RF-CI. RF-CI is the type of RF proposed in Strobl 2007a (“Bias in random forest variable importance measures: Illustrations, sources and a solution”), which I’ve read and commented a little on previously. (See also Strobl 2007b: Unbiased split selection for classification trees based on the Gini Index, which says, “variable selection based on standard impurity measures as the Gini Index is biased. The bias is such that, e.g., splitting variables with a high amount of missing values —- even if missing completely at random (MCAR) —- are artificially preferred.”)

An important difference between CART-like RFs and CI-RFs is that the first builds new n-sample data sets for each tree by sampling with replacement, while the CI-RFs provide each tree with about 63% of the n-sample data set by sampling without replacement. CI-RFs typically grow smaller trees by default, though this is a tweakable parameter (that is, in practice, just like we use CART-like RFs, we customize variants on the original/default CI-RFs).

  • NOTE: in a 2008 follow-up paper (“Conditional variable importance for random forests”), the Strobl group decide that their CI-RFs do not solve all the issues with variable importance, so they also argue that one must use conditional permuation importance as the measure of variable importance; this is also possible to apply to CART-like RFs.
  • GIST: Know that CI-RFs exist, but do not worry about ‘em too much.

Thoughts on VIMPs in a RF

This leads us into variable importances associated with random forests.

For example, the conditional permutation importance advocated by Strobl2008: Is it really “better” than regular permutation importance?

Before reading this paper, I thought the answer was probably “yes,” but now I’m not so sure.

Recall the two causal chains from before:

  1. X2 -> X1 -> Y
  2. X2 <- X1 -> Y

Basically, CPImp acts like the conditional importance described above, associated with a linear model: there is an assumption about the type of causal chain implicit in the method. We may reduce how many variables we look at and look only at direct relationships, but we have lost the true causal story of chain (i). We are telling a partial explanatory story… With regular PImp, the correlated variables would likely be assigned similar importances. Here, we wouldn’t know anything about a direct relationship, yet we wouldn’t be throwing out important variables in the story. The authors bring up LASSO and ElasticNet for comparison. Basically, LASSO chooses one representative from a group of correlated predictors and zeros the rest of them out. This is fine if one’s goal is a parsimonious predictive model, but not fine if one wants to understand the explanatory power of each predictor. ElasticNet was invented to remedy this: it basically forces groups of correlated predictors to share coefficient values (e.g., in group G1, predictors with smaller coefficients are made stronger, while the predictors with larger coefficients are made weaker); now, if one uses an importance measure like standardized, squared coefficients (SSC), all correlated predictors in the same group should share similar importances (similar to PImp). In this method, no important variables are thrown out, but the measures of importance can be considered “biased”.

Written on September 13, 2019