Data Science Masterclass

As A Data Scientist what tips I give to my Younger Version😊

Here are 5 tips.

First and foremost, is data science what you think it is?

9 out of 10 aspiring data scientists I come across equate machine learning with data science. “Data Science” is a loaded, catch-all term. Machine learning is a part of it, but at many major tech companies, product analytics is also an integral part of the data science team. Product analytics is a hidden gem. It is fun but doesn’t get talked about nearly as much. This includes:

A/B test design

Design metrics: Let’s take a video platform as an example. What is the best metric to optimize for that best represents user satisfaction? Should it be number of videos watched? Time spent watching videos? Percent of users that come back in a week to watch another video?

Investigate why metrics change: Why is there suddenly a spike in activity in this cohort of users?

Understand product mechanics: How do button X and feature Y improve the product? Should we redirect page A to B to C or go from A straight to C?

Identify trends and offer strategic suggestions: Argue with data that the company should invest in __ area to stay competitive in the field.

Honestly, when I first started in the field, I was one of those 9 people out of 10 who equated data science with machine learning. Over time though, I’ve come to understand and discover the fun of product analytics, while still getting to work on cool machine learning projects. It was a lucky mistake.

Black boxes are for machine learning, not metrics.

When I first started, I was ready to throw every thing in my tool box left and right: second order differential equations, integrals, derivatives, predictive models. This is classic “when you have a hammer, everything is a nail”. In my experience, great data science work is not synonymous with sophistication. It is rather about asking the right (and usually simple) question about the right data.💦

Particularly for metrics, I strongly advocate for avoiding sophisticated logic flows and hairy prediction models. Let’s return to the video platform example. We can first build a deep neural network (or what have you) to classify whether a video will get attention (say, reach 1K views). Then, we can improve the product to optimize for such videos uploaded, as predicted by this model. I’d suggest directly optimizing for videos that get 1K views—in a well-defined time horizon like a month—rather than optimizing for the output given by this prediction model. Why? Because unless you have extremely high confidence in that model, when that metric drops, the team will spend half of its time scrambling to figure out whether the problem is the model or a real product change. Ideally, top-line metrics should be straightforward. The metric may not be perfect, but at least it is intuitive and free of multiple steps of error propagation. Save machine learning models for in-product use or for inferential understanding purposes instead.

Unlearn academic writing.🤔

Like many people with academic background, I used to communicate my data results like I wrote physics papers for peer-reviewed journals: “With Bonferroni Corrections, the p-value for impressions that result in Y actions, where Y is defined as <greek letter, greek letter, some indicator functions, more symbols, >, is 10E-6, and the change is 24.45968%.”

Scratch that.

Instead: “I recommend launching this feature because it led to a 24% increase in <english> type of behavior.” Place yourself in the reader’s position. The purpose of the result is to make fast, effective business decisions, and therefore being digestible is key. I suggest saving the more rigorous version in a footnote or a separate document for the curious.

Don’t only remain a “numbers” person.🔥

It is easy to have your quantitative skills overshadow product intuition. Don’t constrain your role to spitting out numbers. What does metric X mean for the product, for the business? After all, you are the person who is most familiar with the data set. Help product managers and stakeholders t

924 views10:33