Decision Tree Builder

Interactive Decision Tree Builder – CART Algorithm from Scratch

Build classification trees on your own CSV data and watch the greedy recursive splitting process create interpretable decision rules. Upload a dataset, adjust hyperparameters, and visualize the resulting tree structure with node impurity color-coding and interactive path highlighting.

Technical implementation:

• CART algorithm (Classification and Regression Trees) with greedy top-down induction.

• Dual impurity measures: Gini impurity (1 - Σp²) and Shannon entropy (-Σp·log₂p).

• Numeric feature splits: Evaluates midpoints between sorted unique values, uses cumulative label counting for O(n log n) split finding.

• Categorical feature splits: One-vs-rest binary partitioning for each category value.

• Automatic type inference: Distinguishes numeric vs categorical features based on value distribution (70% numeric threshold).

• Standard stopping criteria: Max depth, min samples per split, min impurity decrease.

Why the implementation matters: Most decision tree tutorials just call sklearn.DecisionTreeClassifier(). This shows you understand:

• Information gain calculation: IG = I(parent) - Σ(n_child/n_parent)·I(child).

• Why midpoint thresholds work for numeric splits.

• How to handle missing values (ignored per split, not imputed).

• The computational cost of exhaustive split search.

Interactive features:

• D3.js tree layout with automatic node positioning.

• Path highlighting: Click "Inspect a row" to see which splits a specific example traverses.

• Node tooltips: Hover to see split rule, information gain, sample counts, class distribution.

• Color-coded impurity: Green (pure) to red (mixed) gradient.

• Export capabilities: SVG for presentations, JSON for model serialization.

• Zoom/pan: Explore large trees interactively.

Dataset flexibility:

• CSV upload or paste.

• Automatic header detection.

• Handles mixed numeric/categorical features.

• Works with multi-class targets.

• Dataset preview table.

Hyperparameter tuning: Adjust max depth, min samples split, and min impurity decrease to observe overfitting vs underfitting trade-offs. The visualization makes it immediately clear when a tree is too shallow (underfits) or too deep (creates single-sample leaves).

Educational value:

• For explaining to stakeholders: "Here's exactly why the model classifies this customer as high-risk" (show the path).

• For debugging: See which features are actually being used (top splits).

• For feature engineering: Notice when a numeric feature gets split at unexpected thresholds.

Comparison to production trees: This implementation uses the same core algorithm as sklearn's DecisionTreeClassifier but with:

• Explicit cumulative counting for numeric splits (not C++ optimized).

• One-vs-rest categorical handling (sklearn uses multi-way splits internally for some tree types).

• Simplified pruning (only pre-pruning via hyperparameters)..