Prediction Problems
A prediction problem asks a Machine Learning model to estimate an unknown outcome using information that is already available.
We know some information, but we don't know the answer yet.
The model learns from previous examples where the outcome was known and then uses what it learned to estimate the outcome for new data.
What Does "Prediction" Mean?
Prediction does not necessarily mean predicting something in the future.
In Machine Learning, prediction can simply mean estimating an unknown value or outcome from known information.
We use available information to estimate a future value.
The house exists, but we may not know its current market value.
The Basic Prediction Structure
Information available to the model.
Uses patterns learned from previous examples.
The outcome we want to estimate.
Example: House Price Prediction
Suppose we want to estimate the price of a house.
1,500 square feet.
A particular neighborhood.
The value we want to estimate.
The model can use the available information and produce an estimated house price.
The Model Learns From Previous Examples
The model does not know the answer automatically. We first provide examples where both the input and the actual outcome are known.
The model uses examples like these to learn a relationship between the available information and the known outcome.
Then We Give It a New Example
After training, we can provide information about a new house where the price is not yet known.
The size is known.
Uses what it learned from previous examples.
Example estimated price.
Prediction Is an Estimate
A prediction should not be treated as a guaranteed answer.
What the model estimated.
What the house actually sold for.
We need appropriate evaluation methods to measure how accurate or useful the model is.
Prediction Can Mean Different Things
The output being predicted depends on the problem.
Predict a numerical value such as price.
Predict which category an example belongs to.
Estimate the likelihood of an outcome.
Prediction Problems Are Everywhere
Estimate how many units customers may buy.
Estimate the likelihood of a particular financial outcome.
Estimate the likelihood of an outcome using available patient information.
Estimate a future temperature using relevant historical information.
What Makes Something a Prediction Problem?
The model receives input data.
We clearly define what we want the model to produce.
Historical examples connect inputs with known outcomes.
Prediction = Estimate an Unknown Outcome From Known Information
A Machine Learning model learns from previous examples and uses the learned relationship to estimate an outcome for new input.
Which Part Is Unknown?
We may know the house characteristics, but the price we want to estimate is unknown.
We may know the customer's behavior, but whether they will purchase is unknown.
We know historical and current information, but tomorrow's temperature is unknown.