Online Robotics Classes for Kids on Sukhumvit, Bangkok

1. Why Split the Dataset

Every supervised ML dataset is divided into training data (often ~80%) for learning patterns and test data (~20%) held out until evaluation. The test set must be isolated from the start so scores reflect real generalization.

2. What Data Leakage Is

Leakage happens when test information mixes into training—for example computing mean and std on the full frame before splitting. The model then “sees” test statistics and scores look artificially good.

Correct order: split first, then fit scaling/imputation only on the training partition and apply those parameters to the test set.

3. Telecom Churn Case Study

1,200 customer records → 80/20 → 960 train, 240 test.

import pandas as pd

df = pd.read_csv("telecom_churn.csv")
print(len(df))  # 1200

random_seed = 42
df_shuffled = df.sample(frac=1, random_state=random_seed).reset_index(drop=True)
train_size = int(0.8 * len(df_shuffled))  # 960

train_df = df_shuffled.iloc[:train_size]
test_df = df_shuffled.iloc[train_size:]

print(len(train_df), len(test_df))  # 960, 240

random_state=42 makes the shuffle reproducible so model comparisons use the same partitions.

4. Verify No Overlap

train_indices = set(train_df.index)
test_indices = set(test_df.index)
intersection = train_indices.intersection(test_indices)
print(intersection)  # set() empty

5. Separate X and y

X_train = train_df.drop(columns=["Churn"])
y_train = train_df["Churn"]
X_test = test_df.drop(columns=["Churn"])
y_test = test_df["Churn"]

print(X_train.shape, y_train.shape)  # (960, 14), (960,)
print(X_test.shape, y_test.shape)    # (240, 14), (240,)

Fourteen input features drive the churn label on both sides of the split.