Fresh rejection writeup — I just did a virtual onsite for a Staff Data Scientist role at Pinterest.
(1) Coding: SQL joining three tables to calculate, for each country, the category with the most impressions, and to calculate the heavy user rate (heavy user defined as active at least 4 days in the past week, with at least 3 features used on one of those days). There was also a simple Python question: find the first n prime numbers.
The interviewer went easy on me this round, I think — really grateful for that.
(2) Hiring manager round: some behavioral questions, but I felt like they were all negative — mostly about what projects had failed, and what I'd improve if I did them again. No positive questions at all. I thought about it for a long time — I don't really have a failed project, so what do you do when you get a question shaped like that? Anyway, the conversation went okay overall.
(3) Stats round:
1 - A billboard question: how do you estimate how many people saw this billboard, and how do you figure out how many impressions it generates (I said split people into three groups — those who live near the billboard, those who pass by it occasionally, and those who never see it — then use a Markov chain to compute a conversion probability).
2 - There's an experiment where an intern forgot to set up the control group — what do you do? What's the difference in emphasis between matching and propensity score weighting? The follow-up was to critique an existing A/B test for pitfalls (I answered with causal inference).
(4) PM round — a senior PM. We talked about projects I'd worked on, then he asked what I actually use Pinterest for (I said searching for cute pet decorations, lol). He asked which phone app I like best, besides Pinterest and my current company's product. I said Xiaohongshu (RedNote), because it has video, photos, and text together — I don't like apps that are just video, haha. Then we spent the rest of the time talking about Xiaohongshu. (Maybe he was doing market research on it?)
(5) DS director round: there's a horizontal recommendation product on the homepage — what metric would you use to tell if it's successful (I said something like ecosystem DAU or time spent). After the experiment launched, the treatment group's homepage CLT dropped, but overall DAU and time spent stayed the same — how would you analyze the possible causes? (I said: check whether there was a major infra launch, whether crash rate went up app-wide, do segmentation to find which segment had the problem, and look at pin quality.) The follow-up was to brainstorm what other segments you could look at.
Discussion
Loading comments…