Meta Data Scientist Interview Experience — Three Rounds Done, Final Technical Session Pushed Two Weeks Out

I finished 3 out of 4 rounds of my virtual onsite yesterday, so while it's still fresh in my memory, let me share my experience. The reason I only finished 3/4 is that after the third round, I found out the recruiter said the final Technical session needs to be rescheduled — it ended up getting pushed two weeks out. Such a slow death (I originally wanted to knock out the whole day in one go, specifically so I could die fast and get reincarnated quickly).

Since I signed an NDA, I won't go into exact detail on the content, but I'll still share how I prepared and roughly what kind of questions I ran into:

Prep process:

1. I'd definitely recommend Jeff Su's videos, but I'll also share advice from a friend of mine who's currently a director at a major retail company, on preparing for Behavioral (the PDF she sent me wouldn't upload, but the content was similar to an article floating around on LinkedIn). The key point: your career story is not your career history. Actually, I made exactly this mistake a while back when I interviewed for a senior manager position at a large multinational HR company (and I bombed it). But that was my first final round, so I'll just chalk it up as practice (sweat).

2. Every session opens with a "short self introduction." So I'd recommend preparing a longer self-introduction (about 2-3 minutes) for the "tell me about yourself" question, and also a short version for icebreakers. Either way, my structure was:

• Who I am right now: besides talking about what I'm currently doing, I'd highlight 2-3 things I'm good at — for example, leading cross-functional teams, analyzing customer insights, being the "firefighter" for ad hoc requests from our VP and SVP whenever they have customer-oriented or purpose-driven questions.
• Who I was and why I became a DS (here I'd quickly connect my PhD experience to data science — not just showing my interdisciplinary background, but also explaining why I became a DS).
• Who I want to be in the future.

3. For Behavioral, I'd strongly recommend finding someone to do mock interviews with you — ideally a manager or someone with more senior experience. If you can't find anyone, you can spend a bit of money online on an interview coach or mock interview. I tried two: Prepfully and IGotAnOffer. Each had its pros and cons. Prepfully was more expensive, and when I did the mock interview with them I hadn't even finished reading the prep material yet, plus I had no confidence at the time, so my mock analytical execution went terribly. But honestly, that experience woke me up quite a bit and changed the direction of my studying afterward — it really did help. As for IGotAnOffer, it was a bit cheaper. I found two coaches, both from different FAANG companies. The first one was upfront that his DS background leaned toward software engineering, so he didn't think he could give me great technical interview advice, but he was a big help on the Behavioral side. The second one happened to be a senior DSA employee at META — besides reviewing my phone screen responses, he also gave me advice on my analytical reasoning process, and finished with a 10-15 minute mock. This part I think helped a lot.

4. ChatGPT is a good friend, but it's not a magic tool: I did use ChatGPT to help polish some of my Behavioral answers when reviewing them, but when you're actually on the battlefield, it's a very different situation. ChatGPT and other AI tools are just aids — the key is still to internalize the situations you might encounter in a Behavioral interview and make them your own. Also, don't make up stories — during the actual interview, the interviewer is very likely to keep digging into your answer, pulling the thread further and further (I'll give an example later).

5. Stories to prepare for Behavioral (if you're interviewing for L5 or above, these are all necessary):

• Any workplace challenge you've encountered and how you ultimately resolved it.
• Any accomplishment at work (big or small) and the impact it ultimately had on the people around you.
• Any conflict you've dealt with.

6. Analytical Execution & Analytical Reasoning: I really recommend Emma Ding's videos — she organizes the material really well. Embarrassing to admit, but I have an MS in Statistics (picked it up along the way during my PhD) with a 3.9 GPA, yet I basically forgot almost everything after graduating (seriously), and honestly I've been bad at combinatorics since high school (I hate that chapter), and any time I run into a probability question I basically just surrender (this is true). This time I worked hard for two weeks to rehab my brain, and looked up a bunch of Bayes' Theorem and Binomial Probability practice problems online. Of course I also reviewed a lot of the required knowledge from a Statistics master's program. Honestly, no matter how the interview turns out, this whole process really helped me going forward (because for the first time in my life I actually understood what combinatorics is doing, lol).

Bayes' Theorem and Binomial Distribution are important, because a lot of the design in social media or e-commerce comes down to click-or-not-click (Binomial), and then when they test probability they layer Bayes' Theorem on top as a combo move. If it's a dice question, that's uniform distribution — also not hard. Once you have the fundamentals down, variations won't throw you off.

7. The following two videos are ones the senior META DSA coach I paid for recommended to me — I personally think they helped a lot:

• When it comes to framing "how to measure success," I think this is a framework that works for both data- and product-oriented roles.
• There's also a good example out there using Facebook Marketplace Metrics.

Behavior Session:

It mainly focused on a few scenarios: (1) Have you ever encountered a coworker feeling "not welcome"? How did you resolve it? (2) What's something you pushed yourself to learn quickly in a short period of time? (3) How do you resolve conflict?

The first type of question was pretty new to me, but I thought it was a great question. The interviewer also worked hard to "dig into" the details — he wanted to know how I handled both sides of the situation (side A disliked B, and B felt unwelcomed by side A). He asked extremely, extremely detailed follow-ups... all I can say is "good thing I actually dealt with something like this before?" (sweat)

Analytical Execution:

The famous fake account problem. Honestly, the descriptions of it on the forum are all so fragmentary that I never actually knew what the question really looked like until I encountered it myself (sweat) — though to be fair, I never bothered to dig through the forum properly either! I'm the type of lazy person who gets tired after scrolling to page two...

But from what I've seen, people who pass or fail this round generally get through 4-5 sub-questions. So if you get the fake account question and manage to get through 4-5 questions in time, that should probably count as passing (?)

My interviewer started with a scenario, asking how to determine whether an account is fake — we discussed this for about five minutes, and I'm not sure if it satisfied him, but then we moved into the actual calculations.

The calculation section is all connected — they tell you the ratio of fake to authentic accounts, then give you scenarios. The first probability question is important because later questions build on its answer. I actually got stuck on the first question for a bit, because I overcomplicated the situation — maybe 30-60 seconds stuck — but once I figured it out, everything after that went smoothly. The first question was actually just P = a / (a+b): once you write out the algebra, it cancels out and gives you the probability p — nothing to do with binomial or Bayes' at all.

The second question involved Binomial. I calculated it a bit quickly at first, so the interviewer asked how I got that number, so I spent another 30 seconds expanding the equation and explaining my reasoning.

If I recall correctly, only the last question used Bayes' Theorem. Honestly, all the numbers in this exam came out clean — for a few of the questions I didn't even need a calculator. On the last question, to repeatedly confirm I hadn't misunderstood the problem, I proactively used "false positive" and "false negative" to check a few numbers (mainly because the numbers came out so clean that I wanted to make sure I hadn't misread the question). The interviewer's reaction seemed positive, so I think in some sense that also confirmed to him that I understood Type I and Type II error.

Just like Behavioral, we left a few minutes at the end to chat with the interviewer, hoping that the minute I got stuck wouldn't be a death sentence for me QAQ

Analytical Reasoning:

For some reason I had this feeling (a mysterious sixth sense?) that this section is really what decides whether you pass or not.

The interviewer for this section never once looked at the whiteboard — he was very engaged, listening and discussing and typing notes at the same time, but he told me "you don't need to use the whiteboard at all! Because I'm not going to look at it anyway" lol

So for this whole section I just looked at the screen and talked with him directly. I got the (in)famous feature recommendation question. It went like this (all very high-level, but you could tell the interviewer wanted me to discuss it from a very top-down, cover-everything perspective):

If you were designing a feature today (basically a new product), how would you convince your boss?

• First is the mission: what is our mission? Growing users? Growing revenue? Or both?
• Once the mission is settled, discuss whether it's actually worth building this new feature/product on top of the existing product. If so, how to explain it using external + internal reasons.
• External means, overall, whether there's really a market opportunity out there. Are there similar products? Who are the competitors? Why should we be the ones carving out this market?
• Internal means, what advantages does adding this feature give us? What can it add?

That's roughly the angle I discussed it from.

If your boss is convinced today, what data would you consider as the basis for analysis?

• Assuming the feature is being added to FB or Instagram itself, what factors would you consider? For example, posts? What kinds of newsfeed clicks? Lifestyle (places traveled, etc.)
• Assuming it's being built on FB, besides FB's own data, you could also think about whether you can pull data from "sibling" products under the same parent, like Instagram, Threads, etc. Usually under the same parent company, certain data can be shared between subsidiaries.
• Is there any other third-party data you could use? Some companies pay for third-party data, and I'd factor that in too.

If you want to measure it, what would you measure?

• You need to clearly define what your north star metric (NSM) is, and what your counter metric (CM) is.
• For example: your NSM might be traffic, and your CM might be conversion rate, because raising traffic doesn't necessarily mean conversion rate also rises.
• Conversion rate can also be broken into several layers depending on the type of engagement — some can be compared test vs. control, some only exist in test — and you need to clearly explain why you'd monitor each of these.
• Maybe because I talk too much(?), I brought up something that fewer people seem to discuss: potential revenue. I said this isn't the NSM or the CM, but it's still a metric worth tracking. Assuming our goal today is monetizing traffic or CTR, then estimating potential revenue matters. If one click brings in $0.5 in revenue, then ultimately, based on the experiment, we can estimate total potential revenue.

Honestly, I only brought this up because of my own personal experience. The interviewer's face lit up a bit at that point, but I couldn't tell whether that "lighting up" meant (a) "I like your answer" or (b) "this rookie is overthinking this, we don't need that." If it's (a), great. If it's (b)... let this be a lesson to everyone reading QAQ

If you needed to run an experiment today, how would you do it? (Since this feature had never launched before, this was a standard A/B test question.)

• First is picking the experiment location. Whenever I said anything, I kept emphasizing "assume we are going to XXXXX" — I think this assumption really matters, since the decision changes depending on the situation.
• Based on my own experience, I'd choose to start with a particular city or metro area (if you have enough money, of course you could pick multiple locations). And that city or metro area should ideally (1) have a large enough population, (2) have a demographic distribution that, in every way, represents the median of the company's product users, or is close to the population distribution of the company's core user base. If the city or metro area you pick instead represents "the population distribution of that whole country," that's also fine, it's just that the result you get will be different — closer to "how people in that country respond to our company's new product" rather than "how our company's current users respond to the new product." (3) Test/control stratification — the interviewer got very detailed here, so I ended up explaining how to build the stratification system, how to score it, how to bucket it into deciles, and then how to split it into test/control from there.

Judging by his expression, he seemed satisfied with my answer on this part? (Of course, if he's a poker face, who knows QAQ)

A bit more I added:

After I finished, for some reason I don't remember, I added something about measuring and deciding at the end of an experiment. I told the interviewer (probably me talking too much again): even if the experiment results come back really good, that doesn't necessarily mean we should launch. In my report to my boss, I'd add cost considerations: technical feasibility, human resources, etc. Because launching a whole new product globally might require investing a lot of money — not just hardware, but software staffing too: do we have enough engineers to build it, do we need to hire more, and how much manpower will maintenance cost afterward — these are all things to consider. Whether it pays back short-term or only long-term, I'd write all of that into the report (and honestly, if I were reporting to a VP or SVP, I would list these as considerations).

After I said this, the interviewer looked pleased, but again... I'm not sure if that was a good sign or if I was just talking too much.

After that was 3-5 minutes of chat time, and we actually ended up going 7 minutes over — hopefully that's a good sign.

Technical Session:

Since the originally scheduled interviewer wasn't available, this got pushed two weeks out QAQ. This is the part I should feel most confident about, but now it just feels like a slow death waiting for it.