Are LinkedIn interviews scripted?
Interviewed with LinkedIn (AI org) last week. I got this weird feeling that their entire interview process is scripted. In the sense that the interviewers have a repository of questions and they choose one of those questions. In fact it seemed like they also have expected answers for those questions. Is this true? If yes, it is a weird way to do an interview. It is actually a test not an interview. I got that this feeling in the middle of my data mining round. The interviewer asked me how to design a recommendation system for users and articles. That itself was weird because my expertise is not recommendation sys but was forced to continue with it. I was ok doing this because one of my close friends has a PhD designing recommender systems at scale and I learnt a bunch of recommender sys design from him. So I brace myself and continue with whatever I know. I mention that one can represent the information you have as a graph / sparse adjacency matrix. Now you can factor the matrix into low-rank components or learn embeddings for users and items separately and use the embeddings to find similar users/items. I decided to go the route of embeddings. I mentioned that the embedding stuff is easier to scale because you can use stochastic gradient descent. In the middle of the interview, the interviewer interrupts me and says: but why are you not using collaborative filtering? I was like what do you mean? This is collaborative filtering. Then I asked her if she wanted me to do the matrix factorization stuff. Then she said yes. That was weird and it triggered alarms that she didn't know what she was asking about. So I compared the low-rank matrix factorization and the embedding approaches and mentioned why the embedding method has advantages w.r.t scaling and integrating new users. I was really surprised because she was at Senior Staff! Then again, in the middle of the conversation she randomly interrupts again and says: your method wont converge. WTF? I asked her why she thought it won't converge. She couldn't say anything except there is not enough data. I gave her two options: you can either reduce the dimension of the embedding or add regularization. She couldn't add anything beyond that to the question. The whole thing seemed really scripted. Then she asked I cannot embed both users and items in the same space. I didn't know. I asked my friend about this later. He mentioned you can do that thats another approach. So this question was totally beyond me. I accept that. Once again: are LinkedIn's interviews scripted? Even if they are scripted, are the questions not matched to the interviewee's expertise? My expertise is search ranking and large scale regression. HC decision: my coding skills are good but my data mining experience needs depth. TC: 290K YOE: 12
Move to the next :)
If I ask a question about recommending articles to users, and the candidate immediately starts talking about matrix factorization/embeddings, that would be a big red flag for me. Specially if he mentions that embeddings can 'scale better because they're trained using SGD'. How do you think matrix factorization is done at scale? SGD. Using fancy methods over simpler ones is a red flag. Infact, a great candidate starts with something like 'lets recommend the most popular articles to everyone' - no personalization. And gradually builds up to a more complex solution. After that and before getting into CF/UserEmbeddings, I'd expect lots of other things to be discussed. Eg: should we featurize articles somehow? I hope this explains why a repository of questions is okay. The interviewer doesn't need to know the latest cutting-edge research in the area - because candidates are expected to start with a simple solution. Exact what a good engineer does in their real job.
Good point
There is nothing fancy wrt embeddings here. In fact it is much simpler to implement than matrix factorization. This was a data mining round. Not a product design round. I agree that I should have started with something much simpler. But the context seemed off because it was a data mining round. Not a system design round.
Yes I got the same impression, only system design was the only that was more authentic, the rest were very shallow and scripted
What position?
Staff ML engg.
I am interviewing for ML eng. Can I pm you with questions about the interview process?
I had the similar experience at LinkedIn the interviewer asked me what is expectation maximization. I told her in layman terms or easy to understand language what it does. After that she asked me to write its mathematical form with full derivation. I told her m not familiar with the derivation as I didn’t review it but can try with whatever I know, the interviewer after kept on asking me about expectation maximiziation and writing full likelihood expressions and derivations. I requested her if she can switch to any other topic in ML that I know with derivations ranging from least squares, ridge regression, svm, logistic regressions, matrix factorization or other optimization techniques as that is my expertise. But interviewer didn’t budge she just wanted me to derive expectation maximization. I got feedback that I failed miserably in that round and hence didn’t make it. Contrast, this with the experience at Snap the interviewer asked me how will you prove a point is from a combination of two distributions, I fumbled and said I don’t know answer right away but I will try. I tried for couple of minutes and he said that he looked at my CV and as I am mostly applied and optimization person so cant answer the question as that’s not my background. He changed question and in the end said I did great on the switched question and I got the offer. LinkedIn interviewer was looking for specific answer and fault is mine as I didn’t review that topic including mathematical derivation. But I would have appreciated if the interviewer could have asked from other area.
I faced this at Google where the interviewer was looking for 'the right answer'
Not scripted. Depends on the quality of the interviewer. It sucks that you got a lazy person. I personally would not ask such a specific question as EM or expect certain solution of the problem. I usually suggest the candidate to pick the algorithm of their choice as a starting point. From there it is easy to have meaningful discussion on all important universal aspects: overfitting, regularization, deployment sgd etc. P.S. designing a recommendation system is a pretty common question, it’s what half of AI at LinkedIn is doing. I would not expect a specific solution. I would prefer starting with something simple and let the candidates go deep into the area that they are comfortable with. Can be feature engineering, or implicit feedback collection or particular modeling approach.
I would also change question if the candidate is not experienced in certain area.
At least for my role, there are a list of "required questions" that it seems an interviewer can choose from. They are looking for particular "signals" (i.e. keywords, concepts) for each of these. Then if there is time after they can go off-script. FWIW I believe the same is true of Google, I got the same 2 questions word-for-word from 2 interviewers there (I alerted the later interviewer that the question had already been asked).
Yes
I had similar experience 6 months ago. There was a lady who has Q/A list, I felt like I was in interrogation in police station. In system design I had similar recommendation related question, how to recommend courses from LinkedIn Learning. Best wishes
Updated the post with HC decision.
I had the same experience a few months ago when a female interviewer pretty much interrogated me instead of interviewing me in system design round. It was as if she was determined to fail me from the minute she walked in. It’s possible that my solution was not perfect but she directed me deeper in that direction for 30 minutes and then kept going on and on about the flaws in it. It was unrecoverable from that point then.