Meta Data Scientist Interview Experience — SQL, Behavioral, and Analytics Execution Rounds

-- Table: active_ads
-- date | ad_id | advertiser_id | creation_source | revenue

-- Table: advertiser_info
-- advertiser_id | advertiser_name | advertiser_country

-- Q1: For different creation sources, calculate daily revenue.
-- Q2: The top ten countries with the least active advertiser users.
-- Q3: For each creation source, what percentage of users spent 1000 more this year than last year.
-- Q4: How would you prove that one source's revenue growth next year is because another source's revenue decreased.

The behavioral round was also some classic questions. For this round I also prepared one example for each question mentioned on the forum — when I got something similar in the actual interview I just tweaked the answer slightly. For the result, it helps to state it clearly, so the interviewer understands exactly how your action solved the problem well and got a good outcome. The questions I got were: give a constructive suggestion, a disagreement, your biggest obstacle, and how you built trust.

The Analytics Execution round is where I gained the most from the forum — thanks to everyone who posted before me — and the interviewer for this round was also really nice and gave a lot of hints. I was asked to draw the distribution of user comments, which is right skewed, and then to mark where the mean, p95, and median would fall, plus roughly what the values would be — worth noting here that the median and p95 could be integers. Then, if you sample all the users to get some subgroups and look at the distribution of the subgroup average comment counts, that's when you need CLT: it becomes a Gaussian distribution, and you're asked again where the mean, p95, and median fall — the mean stays the same, p95 gets smaller, and the median moves closer to the mean. Next there was a product-sense-related question: if you saw the daily average comment count increase, what could be the possible reasons — you could bring up both internal and external reasons. Then: if it's because a lot of fake users joined and left a lot of comments, and you want to build a classifier to catch that, how would you choose your eval metrics. My personal takeaway from this part is that you need to be familiar with a lot of stats and probability theorems, and read a lot of interview experience posts to know which theorem applies in which situation.