This is Part 2 of our Foundations series. In Part 1, we covered data. Now we’re looking at models: what they are, what types exist, and how they represent patterns in data.
What Is a Model?
Think about how a child learns what a dog is. At first, everything with four legs might be “dog.” Over time, they develop a mental model. Dogs come in different sizes. Fur can be long or short. Some have pointed ears, others floppy. Eventually, they can recognize a dog they’ve never seen before, even a strange mixed breed, because their mental model captures the essential patterns of what makes a dog a dog.
A machine learning model does the same thing, but with math instead of intuition. It’s a simplified representation of patterns in your data. A model predicting house prices doesn’t need to know the color of the walls. It captures what matters: how size relates to price, how location affects value, how age impacts worth. Like a globe that doesn’t show every tree but still tells you how continents connect, a model leaves out unnecessary detail and keeps what’s useful.
From Intuition to Math
We build informal models all the time. You know warmer days mean more ice cream sales. You know certain songs fit certain moods. You can catch a ball without calculating trajectories.
Mathematical models make these intuitions precise and scalable. The ice cream relationship becomes a formula that predicts exactly how many more cones you’ll sell when the temperature rises 10 degrees. The ball-catching instinct becomes equations of motion that can guide a robotic arm. Your music preferences become a system that can recommend songs to millions of people simultaneously.
Where it gets interesting is when patterns are too complex for human intuition. Your streaming service processes millions of relationships between shows, viewers, time of day, and viewing habits to find connections no person could spot through observation alone.
Types of Models
Classification models sort things into categories. When a doctor classifies symptoms as one disease or another, when an email filter separates spam from legitimate messages, when satellite imagery labels building damage as minor, moderate, or severe. The model draws mathematical boundaries between groups.
Regression models predict numerical values. How much will this house sell for? How many hospital beds will we need next month? How long until this equipment fails? The model captures how input variables relate to a numerical outcome.
Clustering models find natural groupings without being told what to look for. They’re unsupervised. You give them data and they discover structure: customer segments, disease subtypes, geographic patterns.
Recommendation models match users with content based on patterns of similarity and preference. Netflix, Spotify, Amazon. They work by finding mathematical relationships across millions of users’ behaviors.
Language models represent patterns in how we communicate. Some only analyze text (spam detection, grammar checking). Others both understand and generate text. When ChatGPT writes a response, it’s predicting what should come next based on patterns learned from enormous amounts of text. It understands that “bank” means something different in “river bank” versus “bank account” because of the surrounding context.
Generative models create new content after learning patterns from existing examples. Images, music, text, code. They’ve seen enough examples to produce new ones that are plausible.
Time series models track patterns that unfold over time. Weather forecasting, stock trends, seasonal sales patterns. They capture the rhythm of change and project it forward.
Ensemble models combine multiple models for more reliable results. Like checking several weather apps before planning an outdoor event. No single model is great at everything, but averaging across several tends to smooth out individual mistakes.
Reinforcement learning models learn through trial and error. This is how game-playing AI masters chess and how robots learn to walk. The model tries something, observes the result, adjusts its approach, and repeats.
What We Ask Models to Do
In practice, we use models to:
Predict future outcomes from historical patterns. Sales forecasts, equipment failure timelines, disease progression.
Classify new data into known categories. Medical scans, email filtering, damage assessment.
Optimize decisions across many possible scenarios. Delivery routes, drug discovery, resource allocation.
Detect anomalies by flagging activity that deviates from normal patterns. Fraud detection, infrastructure monitoring, quality control.
Recommend content, products, or connections based on preference patterns across large populations.
The type of model you choose depends on the question you’re asking. We’ll cover this decision-making process in our Mental Models series.
The Bigger Picture
Models are how machines represent what they’ve learned from data. They’re not the data itself, and they’re not the process of learning. They’re the result: a mathematical representation of patterns that can be applied to new situations.
In Part 3, we’ll look at algorithms, the methods machines use to build these models from raw data.
