Machine studying fashions are highly effective, however typically they produce predictions that break human instinct.
Think about this: you’re predicting home costs. A 2,000 sq. ft. house is predicted cheaper than a 1,500 sq. ft. residence. Sounds flawed, proper?
That is the place monotonicity constraints step in. They make sure that fashions comply with the logical enterprise guidelines we anticipate.
Let’s comply with two colleagues, Alan and Aida, on their journey to find why monotonicity issues in machine studying.
The Story: Alan & Aida’s Discovery
Alan is a sensible engineer. Aida is a principled scientist. Collectively, they’re constructing a home value prediction mannequin.
Alan proudly exhibits Aida his mannequin outcomes:
“Look! R² is nice, the error is low. We’re able to deploy!”
Aida takes the mannequin out for testing:
- For a home with 1500 sq ft → Mannequin predicts $300,000
- For a home with 2000 sq ft → Mannequin predicts $280,000 😮
Aida frowns as she appears to be like on the predictions:
“Wait a second… Why is that this 2,000 sq. ft. residence predicted cheaper than a 1,500 sq. ft. residence? That doesn’t make sense.”
Alan shrugs:
“That’s as a result of the mannequin discovered noise within the coaching information. It’s not at all times logical. However, the accuracy is sweet general. Isn’t that sufficient?”
Aida shakes her head:
“Probably not. A reliable mannequin should not solely be correct but in addition comply with logic folks can belief. Prospects received’t belief us if larger properties typically look cheaper. We want a assure. This can be a monotonicity drawback.”
And identical to that, Alan learns his subsequent massive ML lesson: metrics aren’t the whole lot.
What’s Monotonicity in ML?
Aida explains:
“Monotonicity means predictions transfer in a constant path as inputs change. It’s like telling the mannequin: as sq. footage goes up, value ought to by no means go down. We name it Monotone rising. Or, as one other instance, as a home age will get older, predicted costs shouldn’t go up. We name this Monotone reducing.”
Alan concludes that:
“So monotonicity right here issues as a result of it:
- Aligns with enterprise logic, and
- Improves belief & interpretability.”
Aida nodded:
- “Sure, Plus, it helps meet equity & regulatory expectations.”
Visualizing the Drawback
Aida creates a toy dataset in Pandas to indicate the issue:
import pandas as pd
# Instance toy dataset
information = pd.DataFrame({
"sqft": [1200, 1500, 1800, 2000, 2200, 2250],
"predicted_price": [250000, 270000, 260000, 280000, 290000, 285000] # Discover dip at 1800 sqft and 2250 sqft
})
# Type by sqft
data_sorted = information.sort_values("sqft")
# Verify variations in goal
data_sorted["price_diff"] = data_sorted["predicted_price"].diff()
# Discover monotonicity violations (the place value decreases as sqft will increase)
violations = data_sorted[data_sorted["price_diff"] < 0]
print("Monotonicity violations:n", violations)Monotonicity violations:
sqft value price_diff
2 1800 260000 -10000.0
5 2250 285000 -5000.0After which she plots the violations:
import matplotlib.pyplot as plt
plt.determine(figsize=(7,5))
plt.plot(information["sqft"], information["predicted_price"], marker="o", linestyle="-.", coloration="steelblue", label="Predicted Worth")
# Spotlight the dips
for sqft, value, price_diff in violations.values:
plt.scatter(sqft, value, coloration="crimson", zorder=5)
plt.textual content(x=sqft, y=price-3000, s="Dip!", coloration="crimson", ha="heart")
# Labels and title
plt.title("Predicted Home Costs vs. Sq. Footage")
plt.xlabel("Sq. Footage (sqft)")
plt.ylabel("Predicted Worth ($)")
plt.grid(True, linestyle="--", alpha=0.6)
plt.legend()
Aida factors to the Dips: “Right here’s the issue: 1,800 sq. ft. is priced decrease than 1,500 sq. ft. and a couple of,250 sq. ft. is priced decrease than 2,200 sq. ft.”
Fixing It with Monotonicity Constraints in XGBoost
Alan retrains the mannequin and units a monotonic rising constraint on sq. footage and monotonic reducing constraint on home age.
This forces the mannequin to at all times
- improve (or keep the identical) when sq. footage will increase given all different options are fastened.
- lower (or keep the identical) when home age will increase given all different options are fastened.
He makes use of XGBoost that makes it straightforward to implement monotonicity:
import xgboost as xgb
from sklearn.model_selection import train_test_split
df = pd.DataFrame({
"sqft": [1200, 1500, 1800, 2000, 2200],
"house_age": [30, 20, 15, 10, 5],
"value": [250000, 270000, 280000, 320000, 350000]
})
X = df[["sqft", "house_age"]]
y = df["price"]
X_train, X_test, y_train, y_test = train_test_split(X, y,
test_size=0.2, random_state=42)
monotone_constraints = {
"sqft": 1, # Monotone rising
"house_age": -1 # Monotone reducing
}
mannequin = xgb.XGBRegressor(
monotone_constraints=monotone_constraints,
n_estimators=200,
learning_rate=0.1,
max_depth=4,
random_state=42
)
mannequin.match(X_train, y_train)
print(X_test)
print("Predicted value:", mannequin.predict(X_test.values)) sqft house_age
1 1500 20
Predicted value: [250000.84]Alan palms over the brand new mannequin to Aida. “Now the mannequin respects area data. Predictions for bigger homes won’t ever dip under smaller ones.”
Aida exams the mannequin once more:
- 1500 sq ft → $300,000
- 2000 sq ft → $350,000
- 2500 sq ft → $400,000
Now she sees a smoother plot of home costs vs sq. footage.
import matplotlib.pyplot as plt
data2 = pd.DataFrame({
"sqft": [1200, 1500, 1800, 2000, 2200, 2250],
"predicted_price": [250000, 270000, 275000, 280000, 290000, 292000]
})
plt.determine(figsize=(7,5))
plt.plot(data2["sqft"], data2["predicted_price"], marker="o",
linestyle="-.", coloration="inexperienced", label="Predicted Worth")
plt.title("Monotonic Predicted Home Costs vs. Sq. Footage")
plt.xlabel("Sq. Footage (sqft)")
plt.ylabel("Predicted Worth ($)")
plt.grid(True, linestyle="--", alpha=0.6)
plt.legend()
Aida: “Good! When properties are of the identical age, a bigger measurement persistently results in the next or equal value. Conversely, properties of the identical sq. footage will at all times be priced decrease if they’re older.”
Alan: “Sure — we gave the mannequin guardrails that align with area data.”
Actual-World Examples
Alan: what different domains can profit from monotonicity constraints?
Aida: Wherever clients or cash are concerned, monotonicity can affect belief. Some domains the place monotonicity actually issues are:
- Home pricing → Bigger properties shouldn’t be priced decrease.
- Mortgage approvals → Increased revenue shouldn’t cut back approval likelihood.
- Credit score scoring → Longer compensation historical past shouldn’t decrease the rating.
- Buyer lifetime worth (CLV) → Extra purchases shouldn’t decrease CLV predictions.
- Insurance coverage pricing → Extra protection shouldn’t cut back the premium.
Takeaways
- Accuracy alone doesn’t assure trustworthiness.
- Monotonicity ensures predictions align with frequent sense and enterprise guidelines.
- Prospects, regulators, and stakeholders usually tend to settle for and use fashions which can be each correct and logical.
As Aida reminds Alan:
“Make fashions not simply sensible, however wise.”
Closing Ideas
Subsequent time you construct a mannequin, don’t simply ask: How correct is it? Additionally ask: Does it make sense to the individuals who’ll use it?
Monotonicity constraints are one in every of many instruments for designing reliable ML fashions — alongside explainability, equity constraints, and transparency.
. . .
Thanks for studying! I usually share insights on sensible AI/ML methods—let’s join on LinkedIn if you happen to’d wish to proceed the dialog.
