Thread by @svpino

If you want consistent results, you need a consistent process.

Here is every step I go through to tackle new Machine Learning problems.

🧵👇

Step1⃣

▫️What exactly is the problem that you need to solve?
▫️Why do you need to solve this problem?

The answers should give you all the information you need to ensure a solid solution.

👇

Step2⃣

▫️What data do you have access to?
▫️What's the format of that data?
▫️How is that data going to be renewed/expanded?

Then you can focus on cleaning up the data and making it ready to solve the problem.

👇

Step3⃣

▫️How would you solve this problem?
▫️What are some algorithms that you could use?

For each potential solution:

▫️Determine success metric.
▫️Build a quick experiment.

Keep any promising candidate solutions. Discard the rest.

👇

Step4⃣

Iteratively try to improve each candidate's solution, until one surfaces as the best approach.

▫️Pick the best candidate.
▫️Improve its results.

At this point, you don't need the ultimate best solution. You need a working, decent solution.

👇

Step5⃣

▫️Does the solution match the problem?
▫️Is the solution good enough?
▫️Are the trade-offs and limitations acceptable?

Move back one or two steps if the evaluation is not successful.

👇

Step6⃣

▫️How is this solution going to be accessed?

This step is about deploying and operationalizing the Machine Learning solution you just built.

It doesn't matter how good are the predictions of a model if you can't connect them to real data and real users.

🔚