I've been a statistician for two years, and now I'm trying out DS or quant. In late August, Meta's HR reached out to me for an L5 Product Analytics DS interview. After the HR screen, the first technical round was 45 minutes, with 1 SQL question and 1 product question. The setup: an advertiser has a shop on FB and also has their own website. After FB serves a user an ad, they look at whether the customer spends on 'SHOP' or 'WEBSITE'.
SQL first. (1) How does the share of 'SHOP' that advertisers spend perform over the past 30 days? (2) How's FB's model performing? There were two tables, one for revenue and one for number of conversions. I first explained how I'd define 'perform,' then calculated the share based on revenue. For the second question, I only calculated total revenue — the interviewer said I could also think about reshare, and partway through I started talking about product analysis (honestly I was a bit thrown here, because the reshare info wasn't in the two given tables). The interviewer pulled me back to the SQL question. At the end, during the questions round, the interviewer said that for (2) I should think about how to present to a stakeholder, like using a plot. I was even more thrown, since the interviewer hadn't asked about how to present. I felt like I underestimated the SQL — I only went through the interview prep guide once, and I normally use R day to day. I spent 20 minutes on what were two fairly simple technical questions.
The product part asked (1) FB switched to a new ad recommendation algorithm and thinks it's better — what's your take? I clarified whether they wanted a test, and the interviewer said yes. Then I roughly walked through an A/B test framework: units (e.g., why not session-level) --> core + monitoring metrics based on monetization (like weekly/monthly revenue) and user experience, respectively --> randomization assumptions. I paused and asked if that was good so far, the interviewer asked what specific points user experience could include, and I answered by thinking of positive and negative aspects, like time spent + report rate + survey. (2) Then the interviewer asked how to do randomization for this new algorithm where a 50-50 split might not be applicable. I answered that you could randomize users on a rolling basis, since we don't know how big an impact the new algorithm would have (e.g., whether users like it or not, and thus how it affects revenue/ROI). (3) Finally they asked how to recommend to users without running a test. I answered that for existing users, you'd use prior observational data to look at user patterns like shopping behavior, basic characteristics, etc., and build a model. For new users, look at basic characteristics like area and recommend from a general popular list first, and once you collect more info, split them into subgroups/models for personalized recommendation.
Overall it felt like I was the one doing all the talking, the interviewer was expressionless with basically no feedback, and I have no idea whether he was following what I said or had any opinion on it — it felt weird.
Please tear apart my mistakes and shortcomings! Thank you so much!!
Discussion
Loading comments…