Step by Step
S
Select — pick the right algorithm for your problem type
Before touching any data, identify whether your problem is classification, regression, clustering, or something else, since that choice narrows which algorithms are even appropriate.
Example: predicting a continuous house price points you toward regression algorithms; predicting spam vs. not spam points you toward classification algorithms.
P
Prepare — clean, normalize, encode, and split the data
Handle missing values, normalize or standardize numeric features, encode categorical variables, and split the data into training and test sets before any modeling begins.
Example: filling missing age values with the median, converting a "city" column into one-hot encoded columns, and reserving 20% of the data as a test set.
L
Learn — train the model
The chosen algorithm processes the training data and finds the optimal internal parameters (weights, split points, coefficients) that best fit the patterns in that data.
Example: a linear regression model finds the slope and intercept that minimize squared error across the training examples.
I
Interpret — examine what the model actually learned
Look at feature importances, coefficients, or learned splits to understand which factors the model is relying on, and whether that matches domain knowledge and common sense.
Example: checking that a loan-approval model is weighing income and credit history heavily, rather than an irrelevant or inappropriate feature.
T
Test — evaluate on held-out data the model has never seen
Performance on the untouched test set is the only honest measure of how the model will perform on new, real-world data — performance on training data can be misleadingly optimistic.
Example: a model scoring 95% on training data but only 60% on the test set reveals it memorized training quirks rather than learning generalizable patterns.
Applied Walkthrough
1
A student builds a model and reports 98% accuracy — but only measured it on the same data used to train the model.
2
Ask: which step of SPLIT was skipped? The final "Test" step, on truly held-out data never seen during training.
3
Because the reported accuracy came from training data, it tells you almost nothing about real-world performance.
4
The correct fix: go back, hold out a genuine test set from the start, and report accuracy only on that untouched portion.
Exam Application
Exam questions frequently describe a flawed ML workflow (skipping data cleaning, evaluating on training data, not encoding categoricals) and ask you to identify which SPLIT step was done incorrectly or skipped. Knowing the five steps in strict order — Select, Prepare, Learn, Interpret, Test — lets you quickly spot exactly where a described workflow went wrong.
⚠ Common Trap
The most common trap is treating training accuracy as if it were a valid measure of real-world model quality. Training accuracy only tells you how well the model fit the data it already saw — it says nothing about generalization. Only test-set performance, on data the model never touched during training, is an honest measure of quality.
✓ Quick Self-Check
1. What does the S in SPLIT stand for?
Select — choosing the right algorithm for your problem type before doing anything else.
Tap to reveal / hide
2. What happens during the Prepare step?
Cleaning data, handling missing values, encoding categoricals, normalizing/standardizing, and splitting into training and test sets.
Tap to reveal / hide
3. Why is the Test step considered the only honest measure of model quality?
Because it evaluates performance on data the model never saw during training, unlike training accuracy which can be misleadingly high.
Tap to reveal / hide
4. What could explain a model with 98% training accuracy but 55% test accuracy?
The model overfit — it memorized patterns and noise specific to the training data rather than learning generalizable patterns.
Tap to reveal / hide
5. In what order do the five SPLIT steps occur?
Select, Prepare, Learn, Interpret, Test — always in that order.
Tap to reveal / hide