Machine Learning Methods

Three paradigms, one question each

Machine learning finds patterns or relationships in data, and the techniques are grouped into three basic paradigms, each addressing its own problem types: supervised learning, unsupervised learning, and reinforcement learning.

You do not choose between them by preference. The choice follows from the data you have — or could collect — and from the kind of outcome or behavior the system has to learn.

Note that a paradigm is a learning setup rather than a specific algorithm: many algorithms, such as neural networks, can be deployed under different paradigms and on different problem types, and multiple algorithms can address the same problem type.

Supervised learning

A problem is supervised when the data set consists of features or attributes (inputs) that contain target values (outputs). The known answers are what the model learns against.

Two everyday cases show the shape: messages already marked spam or not spam, and houses described by their features together with the price each one actually sold for. In both, someone already knows the answer for the training examples, and the model's job is to reproduce it on new ones.

Supervised learning therefore fits any task where correct outcomes are available and the goal is to predict them for new records.

Classification or regression

Once a problem is supervised, the type of the target decides which of the two supervised problem types you have.

If the target values are categorical (mathematically discrete), the problem is classification:

• Binary classification assigns an individual to one of two predefined and mutually exclusive classes, based on that individual's attributes. A medical diagnosis of disease or no disease from test results is the standard example.
• Multiclass classification assigns an individual to one of several classes. Predicting whether a document is about religion, politics, or finance is an example.

Both are supervised, because the models are trained on examples whose attributes come with correctly labeled objects.

If the target values are mathematically continuous, the problem is regression. Regression estimates the value of a dependent target variable from one or more other variables or attributes that correlate with it — for instance, predicting a house price from the number of bedrooms and bathrooms and the square footage of the house and garden.

The trap in scenario questions is to treat "several categories" as regression or "a numeric value" as classification. Read the target, not the wording around it.

Unsupervised learning

A problem is unsupervised when the data set consists of features or attributes that do not contain labels or target values. The output must then be predicted from the patterns discovered in the input data, and the goal is to discover structure such as groupings within it.

This is the paradigm for exploration: you use it when predefined answers are unavailable, or when the interesting result is the structure itself rather than a value to predict.

Problem types without labels

Several distinct problem types live under this paradigm:

Cluster analysis classifies objects or cases into groups called clusters. It looks for discrete groupings where members of a group are as similar as possible to one another and as different as possible from members of other groups. You define the attributes used to judge similarity, choose a distance function, and specify how many clusters to use.

Anomaly detection identifies rare items, events, or observations that raise suspicion because they differ significantly from the rest of the data — bank fraud and medical errors are the classic uses. Anomalies are also called outliers or novelties.

Reinforcement learning

Reinforcement learning is learning based on interaction with the environment. An agent must learn behavior through trial-and-error interactions with a dynamic environment, and its goal is to maximize the long-term rewards produced by its actions.

Nothing here supplies a correct answer for each situation. The feedback is a reward that may arrive late, and success is judged over a sequence of actions rather than one prediction at a time. Reinforcement learning does not require labeled training data the way supervised learning does.

Rewards are maximized by trading off exploration against exploitation: exploring actions that have uncertain rewards, and exploiting actions that have known rewards. Explore too little and the agent never finds the better path; explore too much and it never cashes in on what it already knows.

A simulated robot choosing a sequence of moves on a changing track is the model case: no state comes with its correct action attached, feedback arrives as rewards, and success depends on a long-term goal.

Choosing between the three

Work from the feedback that the situation actually provides.

1. Do the examples contain the correct targets to predict? Choose supervised learning, then classification for a categorical target and regression for a continuous one.
2. Are there inputs but no targets? Choose unsupervised learning: clustering for groupings, anomaly detection for rare cases, dimension reduction or density estimation for exploration.
3. Must an agent learn a sequence of actions from environmental feedback while optimizing long-term reward? Choose reinforcement learning.

Two scenarios show the sequence at work.

A retailer holds historical examples with a product category and a numeric sale price, and needs to predict the category in one workflow and the price in another. Both targets are known, so both workflows are supervised: the categorical target calls for classification, and the continuous target calls for regression. One dataset, two supervised problem types.

A fraud team holds transaction features but no fraud labels, and wants both natural groupings and rare observations that differ sharply from the rest. With no targets, supervised methods are out; clustering supplies the groupings and anomaly detection flags the rare cases — two unsupervised problem types on the same data.

Key takeaways

• Supervised learning uses target values: categorical targets give classification, binary or multiclass, and continuous targets give regression.
• Unsupervised learning has no labels or targets: its output rests on patterns discovered in the inputs, through clustering, anomaly detection, dimension reduction, or density estimation.
• Reinforcement learning learns from interaction: an agent acts in a dynamic environment by trial and error, trading exploration against exploitation to maximize long-term rewards.
• Choose from the feedback available — known targets, no targets, or environmental rewards — because the paradigm is a learning setup, not a particular algorithm.