Building and Managing a DS Group

Many of the lessons I’ve learned from founding the data science group at AppsFlyer are actually relevant for existing teams as well:
What projects to work on?
What people to hire? When? How many?
How to manage organizational expectations?
How to sell DS to senior management? To your peers?
How much to invest in infrastructure and tools?

One key difference is the startup mentality - you have to win fast. This isn’t the time to build a large infrastructure or take on too risky endeavors. Start small and simple. Make quick wins. The time for larger bets will come later.

In this talk I’ll share some of my own experiences, and show what we can learn from them.
Stay tuned!

Modeling Categorical Variables with Bayesian Networks

Bayesian networks are usually not the first choice for a data scientist trying to model a problem.
However - when encountering multiple categorical variables - they are often the best choice to make.
In this talk I explain what Bayesian networks are, how they operate, and why they succeed where common models - like neural nets or decision trees - fail.

Presented on hayaData 2021: https://www.youtube.com/watch?v=jJXbSy1i-PM

Media Mix Modeling - The Modern Sphinx

Just as the legendary sphinx would ask new riddles each time, media mix modeling presented new challenges in every corner.

Media Mix Modeling is actually an old classic approach for modeling the impact of different ad campaigns on your product purchases. The concept in a nutshell, is to consider the spend in each ad campaign, on a daily basis for a period of say, six months backwards. Then, fit a regression model to explain the overall number of purchases, using the variance in spend of different campaigns.

Sounds easy, right? Sure, but that just doesn’t cut it. One also needs to consider many more effects, such as: ad stocking, diminishing returns, holidays, trend, and seasonality. And that still doesn’t work…

In this talk I present the journey we went through searching for the right modeling approach, the right evaluation approach, some of the lessons we learned, and some of the neatest tricks we used to combine theory and practice.

Geo-Based A/B Testing

When running A/B tests - usually the best thing you can do is to partition your users randomly into subsets A and B, and for a large enough sample the populations will be similar enough by the randomness of the the partition.

However - if we can't partition on a user level, and we must use a geography based partition - the randomness assumption takes a hard hit. For example - if we were to place Tel Aviv in group A and the negev in group B - clearly a hugh bias is inserted into our partition. To avoid this bias we have to carefully design our partition to try and minimize this bias. We also need to analyze our results carefully.

In this talk I present how we design smart geographic partitions, to minimize the inherent bias in the partition. I also discuss how we analyze the results using Google's TBR library.