How Random Forest Hyperparameters Affect Model Performance
On this page
Random Forest performance is driven by three core mechanisms:
- Tree strength (how well each tree fits the data)
- Tree diversity (how different trees are from each other)
- Ensemble averaging (how predictions stabilize across trees)
Each parameter influences one or more of these mechanisms.
Category 1: Tree Structure Parameters (Most Important for Performance)
These parameters control how each individual tree grows and directly affect the bias–variance trade-off.
1. Features per split (mtry / maximum features)
| Low number of features | High number of features |
| Few variables considered at each split | Many variables considered |
| High randomness | Low randomness |
| Trees very different (low correlation) | Trees similar (high correlation) |
| Higher bias | Lower bias |
| Lower variance (better ensemble effect) | Higher variance (less benefit from averaging) |
Interpretation: This parameter controls how similar the trees are to each other. It is the most important tuning parameter because reducing correlation between trees greatly improves ensemble performance.
2. Minimum node size (min.node.size / minimum samples per leaf)
| Small minimum node size | Large minimum node size |
| Very small terminal nodes | Larger terminal nodes |
| Highly complex trees | Simpler trees |
| Captures fine details (including noise) | Produces smoother predictions |
| Lower bias | Higher bias |
| Higher variance (overfitting risk) | Lower variance (underfitting risk) |
Interpretation: This is the primary parameter that controls overfitting. Smaller values allow the model to memorize data; larger values force it to generalize.
3. Maximum tree depth (maximum depth)
| Shallow trees | Deep trees |
| Limited interactions captured | Complex interactions captured |
| Higher bias | Lower bias |
| Lower variance | Higher variance |
| Risk of underfitting | Potential overfitting |
Interpretation: In Random Forest, trees are usually allowed to grow fully because bagging already controls overfitting. Limiting depth is rarely necessary.
4. Minimum samples required to split (minimum samples to split)
| Small minimum split size | Large minimum split size |
| Splits occur easily | Splits occur less often |
| More complex trees | Simpler trees |
| Lower bias | Higher bias |
| Higher variance | Lower variance |
Interpretation: This parameter influences when splitting is allowed, but has less impact than minimum node size.
5. Maximum number of leaf nodes (maximum leaf nodes)
| Few leaf nodes | Many leaf nodes |
| Strong restriction on tree size | Minimal restriction |
| Simpler trees | More complex trees |
| Higher bias | Lower bias |
| Lower variance | Higher variance |
Interpretation: Another way to limit tree complexity. Its role overlaps with the minimum node size.
Category 2: Ensemble Parameters
These parameters control how multiple trees are built and combined.
6. Number of trees (number of estimators)
| Few trees | Many trees |
| Unstable predictions | Stable predictions |
| Higher variance | Lower variance |
| Sensitive to sampling noise | Robust averaging |
| Faster computation | Slower computation |
Interpretation: Increasing the number of trees reduces variance and stabilizes predictions. Performance improves until it plateaus (typically around 300–500 trees).
7. Sample fraction (fraction of data per tree)
| Small sample fraction | Large sample fraction |
| Each tree sees less data | Each tree sees more data |
| Higher diversity between trees | Lower diversity |
| Higher bias | Lower bias |
| Lower variance (better ensemble effect) | Higher variance (trees are more similar) |
Interpretation: Controls how similar trees are. Smaller fractions increase diversity and can reduce overfitting.
8. Bootstrap sampling (sampling with replacement)
| Without replacement | With replacement |
| More unique observations per tree | Some observations repeated |
| Trees more similar | Trees more diverse |
| Slightly higher variance | Lower variance |
Interpretation: Standard Random Forest uses sampling with replacement to increase variability across trees.
9. Class weighting (class weights)
| No class weighting | Balanced class weighting |
| Majority class dominates learning | Minority class emphasized |
| Higher overall accuracy | Better minority detection |
| Lower sensitivity for rare events | Higher sensitivity for rare events |
| Better probability calibration | Possible distortion of predicted probabilities |
Interpretation: This parameter changes how the model prioritizes errors. In clinical prediction, recalibration is often preferred over weighting when probability estimates are important.
Category 3: Feature Handling Parameters
These parameters control how features are evaluated during splitting.
10. Split rule (criterion for split quality)
| Standard split rule (e.g., Gini) | More random split rule (e.g., Extra Trees) |
| Deterministic optimal splits | Randomized split points |
| Lower bias | Slightly higher bias |
| Higher variance | Lower variance |
| Stable performance | More randomness |
Interpretation: The choice of the split rule has a relatively small impact. Standard methods such as Gini impurity work well in most situations.
11. Feature importance method
| Impurity-based importance | Permutation-based importance |
| Fast to compute | Slower to compute |
| Biased toward variables with many categories | Unbiased |
| Less reliable for interpretation | More reliable for interpretation |
Interpretation: This does not affect model performance but is critical for interpreting predictors.
Category 4: Computational / Reproducibility Parameters
These parameters do not affect model performance but ensure reproducibility and efficiency.
12. Random seed (random state)
| Not fixed | Fixed |
| Results vary between runs | Results reproducible |
| Hard to debug | Consistent outputs |
Interpretation: Always fix a random seed to ensure reproducibility.
13. Number of parallel threads (number of jobs)
| Single thread | Multiple threads |
| Slower training | Faster training |
| Lower resource use | Higher CPU usage |
Interpretation: Controls computation speed only.
14. Verbose output (verbosity level)
| Low verbosity | High verbosity |
| Minimal output | Detailed progress logs |
| Cleaner console | More transparency |
Interpretation: Useful for monitoring training progress.
15. Warm start (incremental tree building)
| Disabled | Enabled |
| Model trained from scratch | Trees added incrementally |
| Simpler workflow | Flexible model expansion |
Interpretation: Allows adding more trees without retraining from the beginning.
16. Out-of-bag error estimation (OOB score)
| Disabled | Enabled |
| No internal validation | Built-in validation using unused samples |
| Requires external validation | Quick performance estimate |
Interpretation: Provides an internal estimate of model performance without separate validation data.
Final Conceptual Summary
Random Forest performance depends on balancing:
- Strong trees (deep, small leaf size, high feature usage)
- Diverse trees (low feature usage, smaller sample fraction, bootstrapping)
- Stable averaging (large number of trees)
The most influential parameters are:
- Features per split (controls correlation between trees)
- Minimum node size (controls overfitting)
- Sample fraction (secondary diversity control)
Key Takeaways
- Features per split is the most important parameter because it controls tree correlation
- Minimum node size is the main control for overfitting
- Number of trees stabilizes predictions and should be set sufficiently large
- Sample fraction can further improve diversity between trees
- Most other parameters have smaller or redundant effects