Features, Targets, and Samples

Concept

Supervised learning represents a dataset as samples, features, and a target. A sample is one observation, usually one row. Features are input columns used for prediction. The target, also called a label, is the outcome the model should learn to predict.

Real-World Example

For house-price prediction, each row is one sold house. Size, bedrooms, and age are candidate features; price is the target. The feature matrix is conventionally named capital X, while the target vector is lowercase y.

import pandas as pd

df = pd.DataFrame({
    "size_sqft": [900, 1200, 1600, 2100],
    "bedrooms": [2, 3, 3, 4],
    "age_years": [22, 12, 8, 3],
    "price": [180000, 245000, 310000, 415000],
})

X = df[["size_sqft", "bedrooms", "age_years"]]
y = df["price"]

X is a table with one row per house and one column per feature. y is a single series aligned with those rows. Numerical features are directly numeric; categorical features need later preprocessing before many models can use them.

Reading a Table as a Model Would

The same table has several names for the same basic unit: one row, observation, or sample represents one house. A feature is an input variable such as size_sqft; a target or label is the answer to learn, here price. Capital X conventionally holds the feature matrix because it may contain many columns. Lowercase y holds one target value for each matching sample.

For a churn dataset, tenure_months, monthly_charges, and contract_type could be candidate features, while churn is the target. The target changes with the decision: a spending forecast might use monthly spend as y instead. Defining y first keeps the project focused.

What Belongs in X?

A column is not automatically a good feature. customer_id often identifies a row without describing a useful pattern. A cancellation date recorded after churn must not be used to predict churn because it reveals future information. At prediction time, ask: “Would this value genuinely be available for this new case?” That simple question prevents many leakage mistakes.

Column	Example judgment	Why
`tenure_months`	Good candidate	Available before a churn prediction and may describe behavior.
`customer_id`	Questionable	Usually identifies a row rather than a stable customer pattern.
`account_manager_notes`	Requires review	May be inconsistent or written after risk became known.
`cancellation_date`	Leaky	It is only known after the event being predicted.

An identifier can occasionally be useful for joining or grouping, but that does not make it a sensible numeric predictor. Likewise, a column can exist in a DataFrame and still be invalid, unavailable, or misleading at prediction time.

Interpretation

The same table can have different targets for different decisions. For a churn model, churn might be y; for a spend forecast, monthly spend might be y. The target defines the question, so include it deliberately and never accidentally place it inside X.

Failure Signals

Common Mistakes

Including the target inside the feature matrix.
Using IDs as meaningful numeric measurements.
Including columns created after the predicted event.
Forgetting that X and y rows must stay aligned.

Best Practices

Document row grain, target timing, feature meaning, and known exclusions. Inspect X.shape and y.shape before fitting a model, and choose features based on what would be available in real use.

Interview Perspective

Question: Why are X and y common names? Answer: they conventionally distinguish the matrix of inputs from the vector of outcomes. What the interviewer is testing: whether you understand data representation before discussing algorithms. Follow-up: why might an ID be a poor feature?

Practice Questions

Identify X and y for a customer churn table.
Why is a post-cancellation refund field unsafe for churn prediction?
Write Pandas code to select age and tenure as features.
A table contains customer_id, contract_type, support_calls, churn, and cancellation_reason. Which columns would you exclude from X for a pre-churn prediction, and why?
Why must the rows of X and y stay aligned after filtering a DataFrame?

Quick Quiz

What does one sample usually represent? Answer: one observation or row.
What is the target? Answer: the outcome to predict.
Does a table column always belong in X? Answer: no.

Key Takeaway

Key Takeaways

Samples are rows, X contains appropriate prediction inputs, and y contains the outcome. Good ML starts by defining which information is valid at prediction time.

Next Lesson

Next, use X and y to train a small scikit-learn model end to end.