Types of ML

# ===== The map =====
#
# Supervised: you have inputs X and labels y
#   - Regression: predict a continuous y (price, duration, score)
#   - Classification: predict a discrete y (spam vs not, A/B/C)
#
# Unsupervised: you have only X
#   - Clustering: discover groups (k-means, DBSCAN)
#   - Dimensionality reduction: project to fewer dims (PCA, UMAP)
#   - Density / anomaly: where do points lie? (Isolation Forest, OCSVM)
#
# Self-supervised: labels derived from the data itself
#   - Predicting masked tokens / pixels (BERT, MAE)
#   - Contrastive: same vs different pairs (SimCLR, CLIP)
#
# Reinforcement learning: agent, environment, reward
#   - On-policy (PPO), off-policy (DQN), offline (CQL)

# ===== Picking the right shape =====
# Question:                          Likely shape:
# 'How much will X cost?'            regression
# 'Is this spam?'                    binary classification
# 'Which of 5 buckets?'              multi-class classification
# 'Group similar customers'          clustering
# 'Find weird transactions'          anomaly detection
# 'Compress 200 features to 10'      dimensionality reduction
# 'Choose actions over time'         reinforcement learning

# ===== Same problem, different shapes =====
# 'Predict next-month revenue' -> regression
# 'Predict whether revenue grows next month' -> classification
# The framing changes the algorithm, metric, and evaluation strategy.

# ===== Metrics, picked to match =====
# Regression
#   MAE  (median is robust to outliers)
#   RMSE (penalises large errors more)
#   R^2  (proportion of variance explained)
# Classification
#   Accuracy on BALANCED data
#   F1 / Precision / Recall on imbalanced
#   AUC (ranking quality, threshold-independent)
#   Log loss / Brier for calibrated probabilities
# Clustering
#   Silhouette, Davies-Bouldin, downstream task accuracy
# Anomaly
#   Precision@k (only the top k flagged are inspected)
# RL
#   Cumulative reward, return-per-episode, regret

# ===== Tiny worked examples =====
# Regression (sklearn):
from sklearn.linear_model import Ridge
from sklearn.datasets import fetch_california_housing
from sklearn.model_selection import train_test_split
X, y = fetch_california_housing(return_X_y=True)
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.2, random_state=42)
m = Ridge().fit(Xtr, ytr)
print('R^2:', m.score(Xte, yte))

# Classification:
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_breast_cancer
X, y = load_breast_cancer(return_X_y=True)
Xtr, Xte, ytr, yte = train_test_split(X, y, stratify=y, random_state=42)
m = LogisticRegression(max_iter=2000).fit(Xtr, ytr)
print('acc:', m.score(Xte, yte))

# Clustering:
from sklearn.cluster import KMeans
from sklearn.datasets import make_blobs
X, _ = make_blobs(n_samples=300, centers=4, random_state=42)
km = KMeans(n_clusters=4, n_init='auto').fit(X)
print('inertia:', km.inertia_)

# Dimensionality reduction:
from sklearn.decomposition import PCA
pca = PCA(n_components=2).fit(X)
print('explained var:', pca.explained_variance_ratio_)

# ===== Patterns to internalise =====
# - Frame the question BEFORE picking an algorithm
# - Match metric to outcome: ranking vs absolute prediction vs calibration
# - Try the simplest baseline (Ridge / Logistic / mean predictor) first
# - When a problem looks like 'we need a model', often a rule or threshold ships faster

# ===== Pitfalls =====
# - Using accuracy on heavily imbalanced data -> 99% by predicting majority
# - Clustering 'just because' -> what downstream decision changes?
# - Regression on logged data without exp() back -> reported error is in log space
# - Reinforcement learning when a contextual bandit (or a rule) would do

# Supervised:   classification, regression — labels exist.
# Unsupervised: clustering, dim reduction — no labels.
# Reinforcement: agent learns from rewards.
# Self-supervised: labels come from the data itself (LLMs, MAE).