7. Exercise: Classifying Machine Learning Problems

Type of ML: Supervised Learning
Subtype: Binary classification (buys / doesn't buy)

Justification: You have labeled data (customers with the label "bought" or "didn't buy"). The model learns patterns that distinguish buyers from non-buyers. When a new customer arrives (without a label), the model predicts whether they'll buy or not.

Data needed: Customer history with features (age, previous purchases, browsing, location) and a label (bought yes/no).

Type of ML: Unsupervised Learning
Subtype: Clustering

Justification: You don't have labels (you don't know in advance what segments exist or which customers belong to which segment). The model groups customers by similarity (e.g. K-means). Then you interpret the clusters (e.g. "young customers who buy a lot", "older customers who buy little") and design campaigns per cluster.

Data needed: Customer features (age, purchases, frequency, location) without segment labels.

Type of ML: Supervised Learning
Subtype: Regression

Justification: You have labeled data (houses with features and a known price). The model learns which features influence the price (e.g. more square meters → higher price). When a new house arrives (without a known price), the model predicts how much it should be worth.

Data needed: House features (size, location, rooms, year) and sale price (a numeric label).

Type of ML: It can be Supervised or Unsupervised (depending on the approach)

Approach 1: Supervised (if you have enough labels)
Subtype: Binary classification (fraud / not fraud)
Justification: If you have many transactions labeled as fraud/not-fraud, the model learns fraud patterns and predicts whether a new transaction is fraud.

Approach 2: Unsupervised (if you don't have labels or want to detect unseen anomalies)
Subtype: Anomaly detection (clustering + outliers)
Justification: The model groups normal transactions into clusters; the ones that fall far from the main clusters (outliers) are flagged as suspicious (possible fraud). You don't need labels; the model detects what's "odd".

Most common: A combination of both: supervised for known frauds + unsupervised for new anomalies.

Data needed: Transaction features (amount, location, time, merchant). If supervised: labels (fraud/not-fraud) for training. If unsupervised: only the features.

Type of ML: Reinforcement Learning

Components:

• Agent: The robot
• Environment: Physical space (or a simulation)
• Actions: Moving joints (knee, ankle, hip angles)
• State: The position and velocity of each joint, the position of the center of mass, whether it's on the ground or not
• Reward: +1 for each step forward without falling, -10 if it falls, 0 if it doesn't advance

Justification: There's no labeled data for "the correct way to walk". The robot learns by trial and error: it tries movements, receives feedback (a high reward if it advances without falling, a low reward if it falls), and adjusts its policy to maximize reward.

Data needed: No labeled data; only the simulation (or real environment) where the robot can try actions and receive feedback.

Type of ML: It can be Supervised, Unsupervised or Reinforcement (Netflix uses a combination)

Approach 1: Supervised (rating prediction)
Subtype: Regression (predicting the rating you'd give a movie) or Classification (predicting whether you'll like it or not)
Justification: If you have user ratings for movies (e.g. user X rated movie Y with 4 stars), the model learns to predict what rating you'd give to movies you haven't seen. It recommends the ones with a high predicted rating.

Approach 2: Unsupervised (collaborative filtering, clustering)
Subtype: Clustering of users or movies
Justification: The model groups users with similar tastes or similar movies. It recommends movies that users in your cluster liked, or movies similar to the ones you already watched.

Approach 3: Reinforcement (optimizing long-term engagement)
Subtype: An agent that learns which recommendations maximize engagement (time in the app, movies watched all the way through)
Justification: The agent tries recommendations, receives feedback (if you watched the movie all the way through → a high reward; if you ignored it → a low reward), and adjusts to maximize engagement.

Most common: A combination of supervised (rating prediction) + unsupervised (collaborative filtering) + reinforcement (optimizing long-term engagement).

Data needed: User history (what you watched, what you rated, how long you watched). If supervised: ratings as labels. If unsupervised: only the history. If reinforcement: interaction feedback (whether you watched it all the way through, whether you clicked, etc.).

Type of ML: Supervised Learning
Subtype: Binary classification (spam / not-spam)

Justification: You have labeled data (emails marked as spam or not-spam). The model learns patterns that distinguish spam (e.g. words like "free", "make money", lots of exclamation marks) from not-spam. When a new email arrives, the model predicts whether it's spam.

Data needed: Emails with content (text, subject, sender) and a label (spam / not-spam).

Type of ML: Unsupervised Learning (or related compression techniques)
Subtype: Dimensionality reduction (or autoencoders, a type of neural network)

Justification: There are no labels (there's no "correct answer" for which information to keep; it depends on what you consider "essential"). Techniques like autoencoders learn to compress the image (reduce dimensions) and then reconstruct it while keeping important visual information. The goal is to minimize the difference between the original and reconstructed image.

Data needed: High-resolution images (without labels). The model learns to compress and reconstruct without explicit supervision.

Note: Traditional compression formats (JPEG, PNG) use techniques that aren't necessarily ML; but modern techniques (e.g. autoencoders, VAEs) use unsupervised learning.

Type of ML: Reinforcement Learning

Components:

• Agent: The system that plays chess
• Environment: The chessboard (state = the position of the pieces)
• Actions: Moving a piece (e.g. "move bishop from c4 to f7")
• State: The current position of all the pieces on the board
• Reward: +1 if it wins the game, -1 if it loses, 0 during play (or intermediate rewards if it captures valuable pieces)

Justification: There's no labeled data of "correct moves" for each position (that would be supervised). The agent learns by trial and error: it plays thousands of games (against itself or others), receives feedback (a reward if it wins), and adjusts its policy to maximize victories.

Data needed: No labeled data; only the environment (the chessboard, the rules of the game) where the agent can play and receive feedback.

Note: Modern systems like AlphaZero use RL (self-play: playing millions of games against itself) + search (Monte Carlo Tree Search).

Type of ML: Unsupervised Learning
Subtype: Dimensionality Reduction (e.g. PCA, t-SNE, UMAP)

Justification: There are no labels (there's no "correct answer" for which variables are "most important"; it depends on which patterns you want to see). Algorithms like PCA (Principal Component Analysis) reduce dimensions while keeping the most important variance; t-SNE and UMAP reduce to 2-3 dimensions, optimizing so that similar points end up close together in the plot.

Data needed: A dataset with 100 variables (without labels). The algorithm reduces it to 2-3 variables you can plot.

Common use: Data exploration before training a supervised model; visualizing embeddings (e.g. text or image representations).

Type of ML: Supervised Learning
Subtype: Regression (time series forecasting)

Justification: You have labeled data (past months with known units sold). The model learns temporal patterns (e.g. high sales in December, low in January) and relationships (e.g. promotions → more sales). When next month comes, the model predicts how many units will sell.

Data needed: Sales history (units per month) with explanatory variables (price, promotions, season, location). The label is the number of units sold in each month.

Note: Time series require specific techniques (e.g. ARIMA, LSTM, Prophet) but they're still supervised learning (learning from data with known answers).

Type of ML: It can be Supervised (if you use a pre-trained model) or Unsupervised (if you group without labels)

Approach 1: Supervised with a pre-trained model (more common in production)
Subtype: Multiclass classification (cat, dog, person, landscape, etc.)
Justification: You use a pre-trained model (e.g. ImageNet: 1,000 categories) that already learned to recognize objects. The model classifies each photo (e.g. "cat", "landscape"). That gives you automatic labels without labeling manually.

Approach 2: Unsupervised (visual clustering)
Subtype: Clustering of image embeddings
Justification: You convert each photo into a vector (an embedding) using a neural network; then you group photos with similar embeddings (clustering). The clusters represent visually similar photos. You interpret what each cluster represents (e.g. "beach photos", "pet photos").

Most common: Approach 1 (using a pre-trained model) because it already has useful categories. But if you want to discover your own groupings (not predefined categories), you use approach 2.

Data needed: Photos without labels. If approach 1: a pre-trained model (trained with ImageNet or another labeled dataset). If approach 2: only the photos; the model finds the groupings.

Type of ML: Supervised Learning
Subtype: Binary classification (important / not important)

Justification: You have labeled data (emails marked by the user as important or not). The model learns patterns (e.g. emails from certain senders, with certain words, etc. tend to be important). When a new email arrives, the model predicts whether it will be important.

Data needed: Emails with content (text, sender, subject) and a label (important / not important).

Type of ML: Reinforcement Learning (or optimization algorithms combined with RL)

Components:

• Agent: The system that chooses routes
• Environment: A map of the city with traffic, distances, priorities
• Actions: Choosing the next delivery (which address to go to)
• State: Current location, pending packages, traffic
• Reward: -1 for each minute of time (goal: minimize time); +10 if it meets the delivery priority

Justification: There's no predefined "correct route" (that would be supervised); it depends on dynamic conditions (traffic, urgency). The agent learns by trial and error: it tries routes, receives feedback (a low reward if it takes too long), and adjusts to minimize time.

Data needed: A simulation or real environment where the agent can try routes and receive feedback (time, distance).

Note: In practice, classic optimization algorithms (Traveling Salesman Problem, TSP) are combined with RL to adapt to dynamic conditions.

Type of ML: Supervised Learning (if you have enough labels)
Subtype: Binary classification (genuine / fake)

Justification: If you have reviews labeled as genuine or fake, the model learns patterns (e.g. fake reviews use generic language, have lots of superlatives, the user's profile is recent, etc.). When a new review arrives, the model predicts whether it's genuine or fake.

Data needed: Reviews with content (text, rating, user profile) and a label (genuine / fake).

Alternative approach (unsupervised): If you don't have enough labels, you can use anomaly detection (clustering + outliers): group normal reviews and flag the ones far from the main clusters as suspicious.