Screwing up a machine learning project isn’t difficult

It’s really simple and easy. I will tell you about one of the ways to do it.

First signs related to this appear with judging initial results by people getting involved in a project. Usually, they have a tendency to argue and justify their subjective opinion on the result. It happens because it’s comfortable and easy for them. And it doesn’t require a big effort.

But what’s beneficial for them, it’s bad for the project. Project development based on subjective opinions is a road to nowhere. In this case, you look like running on a hamster wheel without any chance to step further. 

How to cope with it?

1️⃣  If you want to interpret data and results properly, do it based on reliable, objective sources, e.g. experiments or falsifiable hypotheses instead of believing in your assumptions or hand-waving arguments that don’t highlight issues in the big picture. One can easily find hand-waving arguments supporting or contradicting the observations.

2️⃣  If you don't have enough experimental data or it is not possible to design a suitable experiment just take a decision without it, but know that there is a risk associated with the decision. Get back and analyze these decisions later during the project.

3️⃣  Be suspicious.

4️⃣  Think not just one but several steps ahead about the consequences of your decisions and assumptions.

5️⃣  Data or results interpretation isn’t recommended to do alone. Why? Because it's a big chance that you omit something (we’re imperfect human beings with many limitations). 

For years, we at QuantUp devised our own methodology called QuantUp Thinking helping to deal with these challenges of Machine Learning projects. As we don’t believe in existing perfect models, as we're suspicious of every result because it helps us to overthrow questionable points and predict falsifiable hypotheses. 

And we don’t take things for granted. We think more than twice before considering taking something for granted/ verified.

Words by Artur Suchwałko