Overfitting detection for Gradient Boosting — no validation set required
Detect the moment when your model stops learning signal and starts memorizing structure.

---

❓ Why λ-Guard?

In Gradient Boosting, overfitting often appears before the validation error rises.
By that point, the model is already:

• ✂️ Splitting features into extremely fine regions
• 🍃 Fitting leaves supported by very few observations
• 🌪 Sensitive to tiny perturbations

It’s no longer improving predictions, it’s memorizing the training dataset.
λ-Guard detects that moment automatically.

---

🧠 Core Intuition

A boosting model learns two things simultaneously:

Component	Role
Geometry	partitions the feature space
Predictor	assigns values to each region

Overfitting occurs when:

"Geometry keeps growing, but predictor stops extracting real information."

λ-Guard measures three key signals:

• 📦 Capacity → structural complexity
• 🎯 Alignment → extracted signal
• 🌊 Stability → fragility of predictions

---

🧩 Representation Matrix

Every tree divides the feature space into leaves.
We record where each observation falls: Z[i,j] = 1 if sample i falls in leaf j Z[i,j] = 0 otherwise

• Rows → observations
• Columns → leaves across all trees

Think of Z as the representation learned by the ensemble.

• Linear regression → hat matrix H
• Boosting → representation Z

---

📦 Capacity — Structural Complexity

• 🔹 Low C → few effective regions
• 🔹 High C → model fragments space

Late-stage boosting increases C quickly, often without improving predictions.

---

🎯 Alignment — Useful Information

• 🔹 High A → trees add real predictive signal
• 🔹 Low A → trees mostly refine boundaries

"After some trees, alignment saturates."
Boosting continues growing structure even if prediction stops improving.

---

🌊 Stability — Sensitivity to Perturbations

• 🔹 Low S → smooth, robust model
• 🔹 High S → brittle, sensitive model

Stability is the first signal to explode during overfitting.

---

🔥 The Overfitting Index λ

Situation	λ
Compact structure + stable predictions	low
Many regions + weak signal	high
Unstable predictions	very high

Interpretation: measures how much structural complexity is wasted.
Normalized λ ∈ [0,1] can be used to compare models.

🧪 Structural Overfitting Test

Detect if a few training points dominate the model using approximate leverage: H_ii ≈ Σ_trees (learning_rate / leaf_size) T1 = mean(H_ii) # global complexity T2 = max(H_ii)/mean(H_ii) # local memorization

Bootstrap procedure:

1. Repeat B times: resample training data, recompute T1 & T2
2. Compute p-values:
   • p1 = P(T1_boot ≥ T1_obs)
   • p2 = P(T2_boot ≥ T2_obs)

Reject structural stability if:

p1 < α OR p2 < α

---

📊 What λ-Guard Distinguishes

Regime	Meaning
✅ Stable	smooth generalization
📈 Global overfitting	too many effective parameters
⚠️ Local memorization	few points dominate
💥 Extreme	interpolation behavior

---

🧭 When to Use

• Monitor boosting during training
• Hyperparameter tuning
• Small datasets (no validation split)
• Diagnose late-stage performance collapse

---

⚙️ Installation

Install via GitHub:

pip install git+https://github.com/faberBI/lambdaguard.git

from sklearn.ensemble import GradientBoostingRegressor
from lambdaguard.ofi import generalization_index, instability_index, create_model
from lambdaguard.lambdaguard import lambda_guard_test, interpret
from lambdaguard.cusum import lambda_detect
import pandas as pd

# Fit a model
model = GradientBoostingRegressor(n_estimators=50, max_depth=3)
model.fit(X_train, y_train)

# Generalization index
GI, A, C = overfitting_index(model, X_train, y_train)
print('Generalization index: ", GI)

# Lambda-guard test
lg_res = lambda_guard_test(model, X_train)
print(interpret(lg_res))

# CUSUM-based detection
df = pd.DataFrame([
    {"model": "GBR", "n_estimators": 50, "max_depth": 3, "A": 0.8, "OFI_norm": 0.2},
    {"model": "GBR", "n_estimators": 100, "max_depth": 5, "A": 0.85, "OFI_norm": 0.3},
])
cusum_res = lambda_detect(
    df,
    model_name,
    complexity_metric="combined",
    lambda_col="OFI_norm",
    alignment_col="A",
    smooth_window=3,
    cusum_threshold_factor=1.5,
    baseline_points=10
)

📜 Citation

If you use λ-Guard in your research or projects, please cite the following:

Fabrizio Di Sciorio, PhD
Universidad de Almeria — Business and Economics Department

"λ-Guard: Structural Overfitting Detection for Gradient Boosting Models"