File size: 8,290 Bytes
07660e7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 | """
Focused experiment: Training Set Size Sensitivity + Noise Robustness Analysis.
Experiment 1 — Sensitivity Analysis:
How does model performance degrade as we reduce training data?
Trains XGBoost at 10%, 25%, 50%, 75%, 100% of training data
and plots F1 vs. training set size.
Experiment 2 — Noise Robustness:
How robust is the model to label noise (simulating proxy label errors)?
Flips 5%, 10%, 15%, 20% of training labels randomly and measures
degradation in test F1.
Both experiments are well-motivated for this project because:
- Our labels are PROXY labels (heuristic-based), so understanding how
noise in these labels affects model quality is directly relevant.
- Knowing the data efficiency curve tells us whether we need more data
or better features.
"""
import argparse
import os
import warnings
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from sklearn.metrics import f1_score
from build_features import get_feature_columns
from model import ClassicalMLModel
warnings.filterwarnings("ignore")
def run_sensitivity_analysis(
train_feat: pd.DataFrame,
test_feat: pd.DataFrame,
output_dir: str,
fractions: list[float] = None,
n_repeats: int = 3,
):
"""
Train XGBoost at various training set sizes, measure test F1.
Runs multiple seeds per fraction to estimate variance.
"""
if fractions is None:
fractions = [0.1, 0.25, 0.5, 0.75, 1.0]
feat_cols = get_feature_columns(train_feat)
X_test = test_feat[feat_cols]
y_test = test_feat["label"].values
results = []
print("\n" + "═" * 60)
print(" EXPERIMENT 1: Training Set Size Sensitivity")
print("═" * 60)
for frac in fractions:
f1_scores = []
for seed in range(n_repeats):
if frac < 1.0:
sampled = train_feat.sample(frac=frac, random_state=seed)
else:
sampled = train_feat
X_train = sampled[feat_cols]
y_train = sampled["label"].values
model = ClassicalMLModel(params={"n_estimators": 200, "max_depth": 5, "verbosity": 0})
model.fit(X_train, y_train)
y_pred = model.predict(X_test)
f1 = f1_score(y_test, y_pred, average="macro", zero_division=0)
f1_scores.append(f1)
mean_f1 = np.mean(f1_scores)
std_f1 = np.std(f1_scores)
n_samples = int(len(train_feat) * frac)
results.append({
"fraction": frac,
"n_samples": n_samples,
"mean_f1_macro": mean_f1,
"std_f1_macro": std_f1,
})
print(f" {frac*100:5.0f}% ({n_samples:6d} samples) → "
f"F1 = {mean_f1:.4f} ± {std_f1:.4f}")
results_df = pd.DataFrame(results)
# plot
fig, ax = plt.subplots(figsize=(8, 5))
ax.errorbar(
results_df["n_samples"], results_df["mean_f1_macro"],
yerr=results_df["std_f1_macro"],
marker="o", linewidth=2, capsize=5, color="steelblue",
)
ax.set_xlabel("Training Set Size")
ax.set_ylabel("Macro F1 Score")
ax.set_title("Experiment 1: Training Set Size vs. Model Performance")
ax.grid(True, alpha=0.3)
plt.tight_layout()
path = os.path.join(output_dir, "exp_sensitivity_analysis.png")
fig.savefig(path, dpi=150)
plt.close(fig)
print(f" [SAVED] {path}")
csv_path = os.path.join(output_dir, "exp_sensitivity_results.csv")
results_df.to_csv(csv_path, index=False)
print(f" [SAVED] {csv_path}")
return results_df
def run_noise_robustness(
train_feat: pd.DataFrame,
test_feat: pd.DataFrame,
output_dir: str,
noise_rates: list[float] = None,
n_repeats: int = 3,
):
"""
Flip a fraction of training labels to simulate proxy label noise.
Measure how test F1 degrades.
"""
if noise_rates is None:
noise_rates = [0.0, 0.05, 0.10, 0.15, 0.20, 0.25]
feat_cols = get_feature_columns(train_feat)
X_test = test_feat[feat_cols]
y_test = test_feat["label"].values
results = []
print("\n" + "═" * 60)
print(" EXPERIMENT 2: Label Noise Robustness")
print("═" * 60)
for noise in noise_rates:
f1_scores = []
for seed in range(n_repeats):
rng = np.random.RandomState(seed)
y_noisy = train_feat["label"].values.copy()
if noise > 0:
n_flip = int(len(y_noisy) * noise)
flip_idx = rng.choice(len(y_noisy), size=n_flip, replace=False)
y_noisy[flip_idx] = 1 - y_noisy[flip_idx]
X_train = train_feat[feat_cols]
model = ClassicalMLModel(params={"n_estimators": 200, "max_depth": 5, "verbosity": 0})
model.fit(X_train, y_noisy)
y_pred = model.predict(X_test)
f1 = f1_score(y_test, y_pred, average="macro", zero_division=0)
f1_scores.append(f1)
mean_f1 = np.mean(f1_scores)
std_f1 = np.std(f1_scores)
results.append({
"noise_rate": noise,
"mean_f1_macro": mean_f1,
"std_f1_macro": std_f1,
})
print(f" Noise = {noise*100:5.1f}% → F1 = {mean_f1:.4f} ± {std_f1:.4f}")
results_df = pd.DataFrame(results)
# plot
fig, ax = plt.subplots(figsize=(8, 5))
ax.errorbar(
results_df["noise_rate"] * 100, results_df["mean_f1_macro"],
yerr=results_df["std_f1_macro"],
marker="s", linewidth=2, capsize=5, color="coral",
)
ax.set_xlabel("Label Noise Rate (%)")
ax.set_ylabel("Macro F1 Score")
ax.set_title("Experiment 2: Proxy Label Noise vs. Model Performance")
ax.grid(True, alpha=0.3)
plt.tight_layout()
path = os.path.join(output_dir, "exp_noise_robustness.png")
fig.savefig(path, dpi=150)
plt.close(fig)
print(f" [SAVED] {path}")
csv_path = os.path.join(output_dir, "exp_noise_robustness_results.csv")
results_df.to_csv(csv_path, index=False)
print(f" [SAVED] {csv_path}")
return results_df
def main():
parser = argparse.ArgumentParser(description="Run experiments")
parser.add_argument("--data_dir", type=str, default="data/processed")
parser.add_argument("--output_dir", type=str, default="data/outputs")
args = parser.parse_args()
os.makedirs(args.output_dir, exist_ok=True)
train_feat = pd.read_parquet(os.path.join(args.data_dir, "train_features.parquet"))
test_feat = pd.read_parquet(os.path.join(args.data_dir, "test_features.parquet"))
sens_results = run_sensitivity_analysis(train_feat, test_feat, args.output_dir)
noise_results = run_noise_robustness(train_feat, test_feat, args.output_dir)
# combined summary
print("\n" + "═" * 60)
print(" EXPERIMENT SUMMARY & INTERPRETATION")
print("═" * 60)
# interpret sensitivity
min_f1 = sens_results["mean_f1_macro"].iloc[0]
max_f1 = sens_results["mean_f1_macro"].iloc[-1]
improvement = max_f1 - min_f1
print(f"\n Sensitivity: F1 goes from {min_f1:.4f} (10%) to {max_f1:.4f} (100%)")
print(f" Improvement from 10x more data: {improvement:.4f}")
if improvement < 0.02:
print(" → Model is data-efficient; features capture signal well.")
elif improvement < 0.05:
print(" → Moderate benefit from more data; features are decent but more data helps.")
else:
print(" → Strong benefit from more data; consider collecting more labeled traces.")
# interpret noise
clean_f1 = noise_results["mean_f1_macro"].iloc[0]
noisy_f1 = noise_results["mean_f1_macro"].iloc[-1]
degradation = clean_f1 - noisy_f1
print(f"\n Noise Robustness: F1 drops from {clean_f1:.4f} (0%) to {noisy_f1:.4f} (25%)")
print(f" Total degradation: {degradation:.4f}")
if degradation < 0.03:
print(" → Model is highly robust to label noise. Proxy labels are probably fine.")
elif degradation < 0.08:
print(" → Moderate sensitivity to noise. Consider refining proxy labeling heuristics.")
else:
print(" → Significant noise sensitivity. Proxy labels may need manual verification.")
print("\n[DONE] Experiments complete.")
if __name__ == "__main__":
main()
|