Decision trees are a simple but powerful prediction method.
We have seen how a categorical or continuous variable can be predicted from one or more predictor variables using logistic1and linear regression2, respectively. This month we'll look at classification and regression trees (CART), a simple but powerful approach to prediction3. Unlike logistic and linear regression, CART does not develop a prediction equation. Instead, data are partitioned along the predictor axes into subsets with homogeneous values of the dependent variable—a process represented by a decision tree that can be used to make predictions from new observations.
We'll begin with a simple example of one continuous predictor variable, X, and one categorical dependent variable, Y—these variables could be the level of expression of a gene and one of three eye color categories, respectively. Using a sample with n = 60 points equally distributed among three categories (Fig. 1a), let's build a decision tree classifier that predicts the color category Y based on the value of X.
To do this, we'll iteratively split X into intervals that are as homogeneous as possible with respect to Y. In the first iteration, the sample is split into two subsets by considering all possible partitions X < x and X > x, where x is the midpoint between two adjacent observed values of X. We select x to maximize the information gain IG(S1,S2) = I(S) – n1I(S1)/n – n2I(S2)/n, which measures how well the classes are separated by the split of set S into subsets S1 and S2 with n1 and n2 points. By maximizing IG(S1,S2) we favor splits into subsets that are homogeneous with respect to categories and heterogeneous with respect to one another.
The definition of IG(S1,S2) depends on the impurity function I(S), which measures class mixing in a subset. For classification trees, a common impurity metric is the Gini index, Ig(S) = ∑pi(1 – pi), where pi is the fraction of data points of class i in a subset S. The Gini index is minimum (Ig = 0) if the subset comprises a single class and maximum (Ig = (k – 1)/k) when k classes are evenly represented. Figure 1a shows the Gini index for the full sample Ig(S) and for each subset (Ig(S1), Ig(S2)) in three possible splits. For example, for the boundary at X = 20, the pi values for the left subset are (17/20, 3/20, 0) giving Ig(S1) = 0.26 and for the right subset (3/40, 17/40, 20/40) giving Ig(S2) = 0.56. The Gini index information gain is therefore IGg = 0.66 – 20/60 × 0.26 + 40/60 × 0.56 = 0.21. When we calculate IGg for every possible split, we find the maximum IGg = 0.25 at X = 38 (Fig. 1b).
Two other common impurity metrics are entropy, Ie(S) = −∑pilog2(pi) (relatively slow to compute because of the log2), and misclassification error, Ic(S) =1 – max(pi). The information gain functions based on the Gini index and entropy behave similarly—for our sample, both reach a unique maximum, though at slightly different positions (Fig. 1b). The misclassification error is simply the fraction of points in a subset that aren't in the majority vote class. This error is useful for pruning a decision tree but less so for growing one because it is less sensitive to class distribution—for our sample, IGc does not have a unique maximum (Fig. 1b). For example, a sample with class membership (20,20,20) split into (20,10,5) and (0,10,15) gives IGg = 0.13. Splitting instead into (20,20,5) and (0,0,15) gives IGg = 0.22, which is higher and preferred. However, if we used the misclassification error, we would obtain IGc = 0.25 for both splits and could not decide between them. Arguably, the second split is 'purer', since it generates a subset that contains only one class.
Once the first split is chosen, each of the two subsets is split again using the same approach, and the process continues iteratively. The splits generate a decision tree whose nodes correspond to the subsets of the data and whose branches correspond to partitioning of the variable above or below a splitting value for a predictor. Each node is associated with the class that appears most often in the subset, if the cost of misclassification of a data point is independent of class. In Figure 1c we show the full decision tree that classifies our sample based on Gini index—the data are partitioned at X = 20 and 38, and the tree has an accuracy of 50/60 = 83%. Interpreting the decision tree in the context of our biological example, we would associate observations at expression level X < 20 with the green color category. The tree based on entropy (not shown) is similar, with accuracy of 47/60 = 78% and boundaries at X = 25 and 34.
When the dependent variable is continuous, its value can be predicted using regression trees. These predict using the average values of ȳ within each subset, which are selected to minimize the mean square error, MSE = ∑i(ȳ – yi)2/n. For example, the continuous nonlinear function in Figure 2a can be trivially estimated by its average across the full interval, ȳ = 0.58 with MSE = 0.11. We can lower the MSE by partitioning the interval and minimizing n1MSE1/n + n2MSE2/n, the weighted average of MSE of both subsets. Doing this, we find that the boundary at X = 40 yields the lowest value (Fig. 2b), and this split forms the first branching in the regression tree (Fig. 2c). As before, the nodes are subsets but are now associated with the average value of the dependent variable of the points in the subset; 0.75 for X ≤ 40 and 0.23 for X > 40. As we split our regression tree, we get to a state with five leaves and MSE = 0.010 (Fig. 2c).
We can always continue splitting until we build a tree that is 100% accurate, except where points with the same predictors have different classes (e.g., two observations with same gene expression belong to different color categories). However, this would almost always overfit the data (e.g., grow the tree based on noise) and create a classifier that would not generalize well to new data4. To determine whether we should continue splitting, we can use some combination of (i) minimum number of points in a node, (ii) purity or error threshold of a node, or (iii) maximum depth of tree.
For both classification and regression, a useful stopping criterion is to require that each split improves the relative error by at least a, a predetermined value of the complexity parameter. This parameter acts to regularize the cost function of growing the tree5 by balancing the cost with a penalty for adding additional partitions. For example, as we grow our regression tree, we monitor the relative MSE (rMSE) of each split and the amount of decrease a at each split (Fig. 2c). Splitting at X = 49 improves the rMSE by a = 0.05. However, the next candidate split at X = 22 lowers rMSE by only a = 0.007. If we use a cutoff of a = 0.01, this split would not be accepted, and tree growth would end.
An alternative to limiting tree growth is pruning using k-fold cross-validation. First, we build a reference tree on the entire data set and allow this tree to grow as large as possible. Next, we divide the input data set into training and test sets in k different ways to generate different trees. We evaluate each tree on the test set as a function of size, choose the smallest size that meets our requirements and prune the reference tree to this size by sequentially dropping the nodes that contribute least.
In general, trees can be built using forward (tree-growing) or backward (tree-trimming) algorithms. The original CART used tree trimming because the splitting algorithm is greedy and cannot foresee better splits ahead, while trimming grows the whole tree so that the value of the splits can be assessed at the end of the process.
Decision trees can be applied to multiple predictor variables—the process is the same, except at each split we now consider all possible boundaries of all predictors. Figure 3 shows how a decision tree can be used for classification with two predictor variables.
If the data set and the number of predictor variables is large, it's possible to encounter data points that have missing values for some predictor variables. This can be handled by filling in these missing values based on surrogate variables selected to split similarly to the selected predictor.
The creation of the tree can be supplemented using a loss matrix, which defines the cost of misclassification if this varies among classes. For example, in classifying cancer cases it may be more costly to misclassify aggressive tumors as benign than to misclassify slow-growing tumors as aggressive. The penalty is applied as a weight to the impurity index. The node is then assigned to the class that provides the smallest weighted misclassification error. In our example, we did not differentially penalize the classifier for misclassifying specific classes.
Decision trees are very effective and are readily interpreted. However, individual trees can be very sensitive to minor changes in the data, and even better prediction can be achieved by exploiting this variability to grow multiple trees from the same data. This will be the topic of the next column.
Lever, J., Krzywinski, M. & Altman, N. Nat. Methods 13, 541–542 (2016).
Altman, N. & Krzywinski, M. Nat. Methods 12, 999–1000 (2015).
Breiman, L., Friedman, J.H., Olshen, R.A. & Stone, C.J. Classification and Regression Trees (Wadsworth, 1984).
Lever, J., Krzywinski, M. & Altman, N. Nat. Methods 13, 703–704 (2016).
Lever, J., Krzywinski, M. & Altman, N. Nat. Methods 13, 803–804 (2016).
About this article
Cite this article
Krzywinski, M., Altman, N. Classification and regression trees. Nat Methods 14, 757–758 (2017). https://doi.org/10.1038/nmeth.4370
Journal of Ambient Intelligence and Humanized Computing (2020)
Biological and Pharmaceutical Bulletin (2020)
PLOS ONE (2019)
Nature Geoscience (2019)
Parasite Load and Site-Specific Parasite Pressure as Determinants of Immune Indices in Two Sympatric Rodent Species