Random Forest Exercises in R: 20 Real Practice Problems

Explanation: With the formula interface, randomForest() treats a factor response as a classification target and a numeric response as regression. The default mtry for classification is floor(sqrt(p)) where p is the predictor count, so with 4 features it becomes 2. The reported OOB (out-of-bag) error is computed by passing each training row through the trees that did NOT see that row during bootstrap sampling, which is why no separate validation set is needed.

Explanation: The fit object is a regular list, so any of its elements (confusion, err.rate, importance, predicted, votes) can be pulled with $ or [[. The confusion slot is computed from OOB predictions, so it is honest in the sense that it never uses a row to score the same row that trained it. The class.error column equals 1 - diag / rowSums, which is the per-class misclassification rate.

Explanation: OOB error is usually close to test error, but the test set gives a single number to report and to compare against a baseline. predict.randomForest() defaults to type = "response", returning the majority-vote class label. Pass type = "prob" to get the soft scores instead, which is what you need any time you plan to threshold or calibrate.

Explanation: Notice the trap: the headline error rate is 9.5%, which sounds great, but the per-class column shows the model misses 85% of actual frauds. This happens in every imbalanced setup when you optimize plain accuracy: predicting "no" all the time already gets you 90% right. The fix is to change the loss, the threshold, or the sampling, which is what the next exercise tackles.

Explanation: classwt scales how often each class is treated as the "winning" vote when ties are broken at a node, effectively making the trees more willing to predict the rare class. Other levers for class imbalance are sampsize (down-sample the majority within each bootstrap) and cutoff (shift the vote threshold). In production fraud work, threshold shifting plus calibrated probabilities tends to be cleaner than classwt because it leaves the trees untouched.

Explanation: Each row of the probability matrix is the share of trees that voted for that class. Because random forest is itself an ensemble of votes, these scores are not perfectly calibrated, but they are good enough for ranking, lift charts, and AUC. If you need calibrated probabilities for thresholding, fit isotonic regression or Platt scaling on a held-out slice.

Explanation: For regression, the default mtry jumps to floor(p / 3) because trees benefit from more candidate splits when predicting a continuous target. "% Var explained" is 1 - MSE / Var(y), which is an OOB equivalent of R-squared. Note that random forest regression cannot extrapolate beyond the range of the training response: predictions are averages of leaf means, so you will never see a prediction larger than the maximum training y.

Explanation: With only 32 rows, lm() is already overparameterized (10 predictors), so the linear baseline is fragile and RF wins comfortably here. On a real-sized regression problem the gap usually shrinks because linear regression scales gracefully with sample size. The right lesson is: always benchmark RF against a stripped-down lm() before reporting that you used "advanced ML".

Explanation: fit$predicted is the OOB prediction for each training row, so observed - predicted is an honest residual without needing a holdout. A symmetric residual cloud around zero is what you want; heavy tails or a fan shape signals that variance grows with the prediction, which random forest cannot fix on its own. Log-transforming the response, or moving to a quantile regression forest, are the usual next steps.

Explanation: tuneRF() walks outward from the default mtry, multiplying by stepFactor until OOB stops improving by at least improve percent. It is greedier and faster than a full grid search but can miss flat valleys, so for serious tuning prefer caret::train() or a manual grid over mtry and nodesize together. Smaller mtry makes trees more decorrelated; the sweet spot is usually near sqrt(p) for classification and p/3 for regression.

Explanation: The err.rate matrix stores the OOB error at every added tree, so you can see exactly where the curve flattens without refitting four separate models. For most tabular problems, 200 to 500 trees is enough; going to 2000 rarely improves test accuracy and just slows scoring. Use plot(fit) to see the full curve before committing to a final ntree.

Explanation: Default nodesize is 1 for classification (every leaf can be pure) and 5 for regression. Bumping it up acts as a depth regulator and reduces variance, but can also blunt the signal. On a tiny dataset like iris, the OOB curve is almost flat across reasonable values; on noisy data with thousands of rows, the same sweep often shows a clear minimum.

Explanation: expand.grid() is the cheapest way to enumerate a small parameter grid; for larger spaces use caret::train() with trControl = trainControl("oob") so resampling is free. The pattern of OOB across the grid is more useful than the single minimum: if the surface is flat, you are not really tuning, you are just picking a tie. If one corner stands out, retrain with more trees there and confirm before locking it in.

Explanation: Gini importance sums the reduction in node impurity contributed by each variable across all trees, weighted by the rows that reach the split. It is fast (computed during training, no extra passes) but biased toward continuous and high-cardinality predictors. When two variables are correlated, Gini can split the importance arbitrarily between them, so the ranking is more reliable than the magnitudes.

Explanation: Permutation importance shuffles one feature at a time on the OOB rows and measures how much accuracy drops. It is the "fair" version (not biased by cardinality) but is roughly k times more expensive, where k is the number of features. On iris the two rankings happen to agree, but on real datasets with mixed-type features they often disagree, and the permutation ranking is the one to trust.

Explanation: A partial dependence plot fixes every row at the candidate value of one feature, scores the forest, and averages the score across rows. The y-axis for a classification PDP is on a logit-style scale (log(p / (1 - p))), not a probability, which is why values can be negative. PDPs assume features are independent, so they are misleading when predictors are strongly correlated; in that case use ICE (individual conditional expectation) plots or SHAP values instead.

Explanation: ranger is API-compatible enough that switching from randomForest is usually a name swap (ntree becomes num.trees, mtry is the same), but it is dramatically faster on data with more than a few thousand rows and parallelizes across cores by default. The OOB error matches randomForest here because the algorithm is the same; only the implementation differs.

Explanation: Two forests trained with the same hyperparameters won't be byte-identical (their RNG streams differ) but should agree on the vast majority of rows on a clean dataset. A drop below 95% usually points to a column-order or factor-level mismatch, not a real algorithmic disagreement. The predict.ranger() return is a list with a $predictions slot, which is the common gotcha when porting code.

Explanation: Stratified sampling preserves the marginal distribution of the response, which matters more when the class ratios are skewed. A single split is still a noisy estimate; for a more stable number, repeat the split (or use k-fold CV) and average the per-fold accuracies. Random forest's OOB error usually tracks the test error closely, so on small data the OOB number is fine to report.

Explanation: Recall (or sensitivity, or TPR) is the right per-class metric whenever missing a true positive is costly: disease screening, fraud, defect detection. Precision (diag / colSums) is its mirror image and matters when false alarms are costly. The full picture is best read off the confusion matrix itself; single-number metrics like accuracy or F1 are convenient but always discard structure.