Can the same dataset be used for both regression and classification?

What you need to know

The target defines the task

The same rows of features can support many targets:

One dataset, two models

1import numpy as np2import pandas as pd3from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier45rng = np.random.default_rng(42)6n = 10007df = pd.DataFrame({"area": rng.uniform(500, 2500, n), "dist_km": rng.uniform(2, 30, n)})8df["price"] = 20 + 0.07 * df.area - 1.2 * df.dist_km + rng.normal(0, 10, n)   # lakh910# Same rows, two different targets11df["band"] = pd.cut(df.price, bins=[0, 60, 120, np.inf], labels=["budget", "mid", "premium"])12X = df[["area", "dist_km"]]1314reg = RandomForestRegressor(random_state=0).fit(X, df.price)   # target: a number15clf = RandomForestClassifier(random_state=0).fit(X, df.band)   # target: a category1617flat = pd.DataFrame({"area": [1150], "dist_km": [12]})18print("regression    :", reg.predict(flat).round(1), "lakh")19print("classification:", clf.predict(flat), clf.predict_proba(flat).round(2))20print("class order   :", clf.classes_)

regression    : [82.6] lakhclassification: ['mid'] [[0.03 0.97 0.  ]]class order   : ['budget' 'mid' 'premium']

Same flat, same features. The regressor says 82.6 lakh. The classifier says "mid", with 97% confidence. pd.cut did the conversion from number to band.

What you lose when you bucket

The "mid" band runs from 60 to 120 lakh. A flat at 61 lakh and one at 119 lakh get the same label, and a flat at 59 lakh looks as different from the 61-lakh flat as a flat at 20 lakh does. The model can no longer learn that 61 is close to 59. So bucketing usually loses accuracy.

It is still the right call when:

• The decision really is categorical: approve or reject, send to manual review or not.
• The exact number is noisy or not trusted, but the band is.
• The users of the prediction act on bands, like "show premium listings to this user".

A useful middle path: predict the number, then bucket the prediction. You keep the full information in training and can move the cut-offs later without retraining.

Going from classes to numbers

Coding classes as 0, 1, 2 and running regression only makes sense when they are ordered and evenly spaced, such as a 1-to-5 star rating. Coding "Billing = 0, Network = 1, Sales = 2" as numbers invents an order that does not exist.

A real-life example

A lender first built a "default: yes or no" classifier. The finance team then asked, "How much will we lose?" A borrower who defaults after paying 11 of 12 EMIs costs far less than one who never pays. The team kept the classifier for approval decisions and added a regression model for the amount lost given default, trained on the same applications. The combination let them price loans by risk.

Follow-up questions to expect

• "Is predicting a star rating regression or classification?" — Either. Ratings are ordered, so regression (or ordinal regression) uses that order; classification ignores it. Try both and compare on the metric that matters.
• "How would you pick the band cut-offs?" — From the business, such as the price ranges the sales team already uses, not from what makes the model score higher.
• "Can one model do both?" — Yes, a neural network can have two heads, one for each output, trained together.