File size: 10,166 Bytes
0ed43fe | 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 | """
EcoCart Customer Segmentation β Bias Detection & Mitigation
Task 2 β Demonstrates urban-rural bias in K-Means segmentation and
applies reweighing to fix it.
NCI MSCAI | Fundamentals of AI TABA 2026
Run: python3 task2_segmentation.py
Out: bias_before_after.png, disparate_impact.png
"""
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.cluster import KMeans
from sklearn.preprocessing import StandardScaler
RNG = np.random.default_rng(42)
# ββ 1. Generate biased customer data ββββββββββββββββββββββββ
# Urban customers have more data, higher frequency, higher spend β mimicking
# a real scenario where the platform launched in cities first.
def generate_biased_data(n_urban=300, n_rural=100):
# Urban: higher frequency and spend on average
urban = pd.DataFrame({
"freq": RNG.normal(6.0, 2.0, n_urban).clip(0.5),
"spend": RNG.normal(120, 40, n_urban).clip(10),
"recency": RNG.exponential(10, n_urban).clip(1, 90),
"region": "urban",
})
# Rural: lower frequency and spend (platform is newer there)
rural = pd.DataFrame({
"freq": RNG.normal(3.0, 1.5, n_rural).clip(0.5),
"spend": RNG.normal(65, 30, n_rural).clip(10),
"recency": RNG.exponential(15, n_rural).clip(1, 90),
"region": "rural",
})
df = pd.concat([urban, rural], ignore_index=True)
df["freq"] = df["freq"].round(1)
df["spend"] = df["spend"].round(0)
df["recency"] = df["recency"].round(0)
return df
# ββ 2. Segment with K-Means ββββββββββββββββββββββββββββββββ
def segment(df, features=["freq", "spend", "recency"]):
scaler = StandardScaler()
X = scaler.fit_transform(df[features])
km = KMeans(n_clusters=3, random_state=42, n_init=10)
df = df.copy()
df["cluster"] = km.fit_predict(X)
# Label clusters by mean spend (High/Medium/Low)
means = df.groupby("cluster")["spend"].mean().sort_values(ascending=False)
label_map = {means.index[0]: "High Value",
means.index[1]: "Medium",
means.index[2]: "Low Value"}
df["segment"] = df["cluster"].map(label_map)
return df
# ββ 3. Bias metrics ββββββββββββββββββββββββββββββββββββββββ
def compute_fairness(df):
urban = df[df.region == "urban"]
rural = df[df.region == "rural"]
u_high = (urban.segment == "High Value").mean()
r_high = (rural.segment == "High Value").mean()
di = r_high / u_high if u_high > 0 else 0
return {
"urban_high_pct": round(u_high * 100, 1),
"rural_high_pct": round(r_high * 100, 1),
"disparate_impact": round(di, 3),
"fair": di >= 0.8,
}
# ββ 4. Mitigation: reweigh + balanced re-sample ββββββββββββ
def mitigate(df):
"""
Fix 1: Balance the dataset by oversampling rural customers.
Fix 2: Add a 'distance_adjusted_spend' feature that normalises
spend by delivery cost (rural customers pay more for delivery,
so their raw spend understates their purchase intent).
Fix 3: Post-processing β reassign borderline rural customers using
a lowered threshold derived from the rural spend distribution.
"""
df = df.copy()
# Oversample rural to match urban count
rural = df[df.region == "rural"]
urban = df[df.region == "urban"]
rural_up = rural.sample(n=len(urban), replace=True, random_state=42)
balanced = pd.concat([urban, rural_up], ignore_index=True)
# Adjust spend: rural delivery costs ~β¬12 more on average
balanced["adj_spend"] = balanced.apply(
lambda r: r["spend"] + 12 if r["region"] == "rural" else r["spend"],
axis=1,
)
# Adjust frequency: rural customers batch orders
balanced["adj_freq"] = balanced.apply(
lambda r: r["freq"] * 1.5 if r["region"] == "rural" else r["freq"],
axis=1,
)
# Re-segment on adjusted features
scaler = StandardScaler()
X = scaler.fit_transform(balanced[["adj_freq", "adj_spend", "recency"]])
km = KMeans(n_clusters=3, random_state=42, n_init=10)
balanced["cluster"] = km.fit_predict(X)
means = balanced.groupby("cluster")["adj_spend"].mean().sort_values(ascending=False)
label_map = {means.index[0]: "High Value",
means.index[1]: "Medium",
means.index[2]: "Low Value"}
balanced["segment"] = balanced["cluster"].map(label_map)
# Post-processing: promote top rural "Medium" and "Low Value" customers
# to "High Value" until disparate impact reaches 0.85 (above 0.8 threshold)
rural_mask = balanced.region == "rural"
urban_mask = balanced.region == "urban"
urban_high_rate = (balanced[urban_mask].segment == "High Value").mean()
target_rate = urban_high_rate * 0.85
n_rural = rural_mask.sum()
target_rural_high = int(target_rate * n_rural)
current_rural_high = ((balanced[rural_mask].segment == "High Value")).sum()
need = target_rural_high - current_rural_high
if need > 0:
# Promote from Medium first, then Low Value
candidates = balanced[rural_mask & (balanced.segment != "High Value")]
if len(candidates) > 0:
promote = candidates.nlargest(min(need, len(candidates)), "adj_spend").index
balanced.loc[promote, "segment"] = "High Value"
return balanced
# ββ 5. Plots ββββββββββββββββββββββββββββββββββββββββββββββββ
SEG_COLORS = {"High Value": "#10b981", "Medium": "#f59e0b", "Low Value": "#ef4444"}
def plot_before_after(before_df, after_df, before_fair, after_fair):
fig, axes = plt.subplots(1, 2, figsize=(14, 5.5))
fig.patch.set_facecolor("#0d1117")
for ax, df, fair, title in [
(axes[0], before_df, before_fair, "BEFORE mitigation (biased)"),
(axes[1], after_df, after_fair, "AFTER mitigation (reweighed + adjusted)"),
]:
ax.set_facecolor("#0d1117")
for seg in ["High Value", "Medium", "Low Value"]:
mask = df.segment == seg
for region, marker in [("urban", "o"), ("rural", "^")]:
rmask = mask & (df.region == region)
ax.scatter(df.loc[rmask, "freq"], df.loc[rmask, "spend"],
c=SEG_COLORS[seg], marker=marker, s=25, alpha=0.6,
label=f"{seg} ({region})" if ax == axes[0] else None)
di = fair["disparate_impact"]
color = "#ef4444" if not fair["fair"] else "#10b981"
ax.set_title(f"{title}\nDI = {di:.3f} {'β BIASED' if not fair['fair'] else 'β FAIR'}",
color="white", fontsize=11)
ax.set_xlabel("Purchase frequency / month", color="white")
ax.set_ylabel("Avg spend (β¬)", color="white")
ax.tick_params(colors="white")
ax.grid(True, alpha=0.1, color="white")
axes[0].legend(fontsize=7, facecolor="#0d1117", edgecolor="#334155",
labelcolor="white", loc="upper right", ncol=2)
plt.tight_layout()
plt.savefig("output/bias_before_after.png", dpi=150,
bbox_inches="tight", facecolor="#0d1117")
plt.close()
def plot_di(before_fair, after_fair):
fig, ax = plt.subplots(figsize=(8, 4))
fig.patch.set_facecolor("#0d1117")
ax.set_facecolor("#0d1117")
cats = ["Urban β High", "Rural β High", "Disparate Impact"]
before_vals = [before_fair["urban_high_pct"], before_fair["rural_high_pct"],
before_fair["disparate_impact"] * 100]
after_vals = [after_fair["urban_high_pct"], after_fair["rural_high_pct"],
after_fair["disparate_impact"] * 100]
x = range(len(cats))
w = 0.35
ax.bar([i - w/2 for i in x], before_vals, w, label="Before", color="#ef4444", alpha=0.85)
ax.bar([i + w/2 for i in x], after_vals, w, label="After", color="#10b981", alpha=0.85)
ax.axhline(80, color="#fbbf24", linewidth=1.5, linestyle="--", label="DI threshold (80%)")
ax.set_xticks(x)
ax.set_xticklabels(cats, color="white")
ax.set_ylabel("Percentage", color="white")
ax.set_title("Fairness metrics before vs after mitigation", color="white", fontsize=12)
ax.tick_params(colors="white")
ax.legend(fontsize=9, facecolor="#0d1117", edgecolor="#334155", labelcolor="white")
ax.grid(True, axis="y", alpha=0.15, color="white")
plt.tight_layout()
plt.savefig("output/disparate_impact.png", dpi=150,
bbox_inches="tight", facecolor="#0d1117")
plt.close()
# ββ 6. Main βββββββββββββββββββββββββββββββββββββββββββββββββ
def main():
print("="*70)
print("EcoCart Customer Segmentation β Bias Detection & Mitigation")
print("="*70)
# Generate and segment (biased)
df = generate_biased_data()
df = segment(df)
before = compute_fairness(df)
print(f"\nBEFORE mitigation:")
print(f" Urban -> High Value: {before['urban_high_pct']}%")
print(f" Rural -> High Value: {before['rural_high_pct']}%")
print(f" Disparate Impact: {before['disparate_impact']}")
print(f" Fair (DI >= 0.8)? {before['fair']}")
print(f"\n Segment counts:")
ct = df.groupby(["region", "segment"]).size().unstack(fill_value=0)
print(ct.to_string(index=True))
# Mitigate
fixed = mitigate(df)
after = compute_fairness(fixed)
print(f"\nAFTER mitigation:")
print(f" Urban -> High Value: {after['urban_high_pct']}%")
print(f" Rural -> High Value: {after['rural_high_pct']}%")
print(f" Disparate Impact: {after['disparate_impact']}")
print(f" Fair (DI >= 0.8)? {after['fair']}")
# Plots
plot_before_after(df, fixed, before, after)
plot_di(before, after)
print("\nWrote: bias_before_after.png, disparate_impact.png")
if __name__ == "__main__":
main()
|