Real-World Problem → ML Problem
A real-world problem usually starts as a business, product, or practical question. Before a model can solve it, we need to translate that question into something that can be represented using data.
Don't ask the model to solve a vague problem. Define exactly what information it receives and what result you want.
Turning a real-world problem into an ML problem is mainly about defining the input, the desired output, and the examples from which the system can learn.
Start With a Real-World Question
Imagine an online store wants to increase sales. Someone might ask:
This is a useful business question, but it is not yet a clearly defined Machine Learning problem.
Define What We Want to Predict
We first need to decide what the model should produce.
Now the problem has a specific target: determine whether a customer is likely to purchase the product.
Identify the Information Available
Next, we need to determine what information is available about each customer.
What the customer purchased in the past.
Pages viewed and products interacted with.
Products added to the shopping cart.
The model cannot use information that we do not provide as input.
Find Historical Examples
If we want the system to learn from past customer behavior, we need historical examples.
The historical examples give us both information about the customers and the outcome that actually happened.
Now the ML Problem Is Clear
Previous purchases, activity, cart behavior, and other available information.
Find patterns that relate customer information to purchasing behavior.
Whether the customer is likely to purchase.
Another Example: House Prices
The same conversion works for house-price prediction.
The Real-World Problem Is Usually Bigger
A real business problem often contains many questions, requirements, and constraints.
We need to narrow it down to one clearly defined prediction or decision.
This is a business goal, not a specific ML prediction.
Now we have a defined output that a model can attempt to predict.
Don't Force ML Into a Simple Rule
Not every real-world problem should become a Machine Learning problem.
If the account balance is zero, prevent a withdrawal.
Predict whether a transaction is likely to be fraudulent based on historical examples.
A Simple Conversion Process
What are we trying to achieve?
What information do we have?
What do we want to predict?
Clearly defined prediction or decision.
Problem Definition Comes Before Model Selection
A good Machine Learning project starts by defining the problem clearly. If the problem itself is vague, choosing a sophisticated algorithm will not fix it.
Turn These Into ML Problems
Possible ML problem: predict whether a customer is likely to leave.
Possible ML problem: predict whether a transaction is fraudulent.
Possible ML problem: predict tomorrow's temperature using available historical and relevant weather information.