Machine Learning Basics

Overview

Machine learning (ML) teaches computers to improve at a task through experience (data) instead of hand-written rules. You provide examples; an algorithm adjusts internal parameters to minimize error on those examples, then generalizes to unseen inputs.

Two dominant paradigms:

• Supervised learning: each example has a label (spam/not spam, price, category). The model learns input → label mappings.
• Unsupervised learning: no labels. The model finds structure—clusters, anomalies, or compressed representations.

Training fits the model; inference applies it to new data in production.

Syntax / Usage

Typical supervised workflow:

1. Collect & clean data
2. Split: train / validation / test
3. Choose model & loss function
4. Train (optimize weights on train set)
5. Tune hyperparameters using validation set
6. Evaluate once on held-out test set
7. Deploy for inference

Common supervised tasks:

Unsupervised examples: customer segmentation (clustering), dimensionality reduction (visualization), anomaly detection (unusual server metrics).

Libraries: scikit-learn (tabular ML), PyTorch / TensorFlow (deep learning), Hugging Face (pretrained models).

Examples

Training a simple classifier with scikit-learn:

from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score

X_train, X_test, y_train, y_test = train_test_split(features, labels, test_size=0.2)
model = LogisticRegression()
model.fit(X_train, y_train)

predictions = model.predict(X_test)
print(accuracy_score(y_test, predictions))

Inference in an API route:

def predict_churn(user_features: list[float]) -> float:
    probability = model.predict_proba([user_features])[0][1]
    return float(probability)

Common Mistakes

• Evaluating on training data and claiming high accuracy (overfitting)
• Leaking test-set information into feature engineering or tuning
• Using accuracy alone on imbalanced datasets (99% non-fraud → useless model)
• Deploying without monitoring for data drift (inputs change over time)
• Retraining on production logs without privacy review or label quality checks

See Also

ai-fundamentals neural-networks embeddings