LinkedIn Machine Learning Mock Interview - Design a recommendation engine

[@praveen_me]

This is fantastic. The idea of getting to a simple baseline model then starting to increase complexity is something many people miss.

[@shantanubapat6937]

The interviewee would fail if this was an actual interview. He entirely missed the core question: How to recommend new job that no user has intereacted before. This is a cold start problem and collabrative filtering cannot address it.

[@venkateshgunda25]

Exactly. SOOO much talk about RecSys theory that nobody asked. The Cold-start problem needs to be an Item-based Filtering, and then once it is warm and a certain amount of critical interactions happen, adjusting the weightage to a collab based filtering would have worked great. Basically, Hybrid Collab filtering.

[@harishgona]

Yes I was expecting he would eventually get to the content based approach 🤷‍♂️

[@iqjayfeng]

Thanks for watching! Get ready to ace any question thrown your way during the interview. You can find our LinkedIn Machine Learning Engineer interview guide here: www.interviewquery.com/interview-guides/linkedin-machine-learning-engineer

[@LCRedemption]

Ved should probably consider to become a politician, who usually talks a shit tons of useless nonsense before hitting the point.

[@be051268]

haha

[@AkhilVerghese]

This is pretty good, but as someone who does conduct ML interviews there are a few areas I would have pushed harder: 1) A matrix factorization method is great in a system where new items aren't particularly important (or there are a lot of users that will click on new things quickly). But if the features are completely generated based on commonality of interactions, what happens to new jobs that have no interactions? Usually there are at least some categorisations that need to be attached (based on embeddings in the item, company name or company category, or tags by the poster) that can help with initial recommendation (until enough people apply and matrix factorization can take over). 2) Didn't really answer the question on what the success metrics would be to compare in AB testing or if you don't have an existing solution. I have no doubt Ved could answer these questions, I just feel a few more followups would be more realistic.

[@iqjayfeng]

Good points Akhil - are you interested in doing a mock interview btw on the channel?

[@MohamedBoukhari-go4yv]

That's what i thought too. I was wondering for a new job where we have no interaction, how to get the feature vector to clusterize then ? so you are saying that using more information about the job like company name etc would help get that feature vector ? or am i missing something.

[@sakshamjindal6992]

What do you say when you don't know shit about the interview question ? "That limits our option, but not our motivation and excitement to do this."

[@danishshaikh2952]

lol 😂

[@prashantmaheshwari4066]

No interviewer is this much agreeable. No cross-questioning, allowing to present a tangential answer for more than half the time. A mock should be mock where the chain of thoughts is broken due to questions and you have to continue/connect the answer from where you left.

[@fahnub]

true

[@jaad9848]

This is something that is true of all these paid/volunteered interviews on YT. Its why these mocks are unrealistic

[@kunalsaini1805]

This is great, shedding light on every aspect, from algorithm to productionalization to how to proceed with the test and roll it out at the end

[@The_Pavanputra]

I watch this video before my every machine learning engineering interview and believe me I ace ML part of it everytime. Thanks for this gem.

[@MrTupapi0826]

What other parts are there? If you don’t mind sharing.

[@YG-hu2cb]

Collaborative filtering is famous for not able to generalize to new users/items. The interviewer explicitly asked how to recommend similar jobs to the new jobs at the beginning. What's the candidate's answer to that??

[@sakshamjindal6992]

And he is doing a PhD

[@vandana9089]

Matrix factorization can be used to find similarities between items that already have interaction data. New items will not have interaction data, so it can't be applied to the question asked. Its an interesting discussion regardless.

[@shairuno]

Definitely he has an experience, not just technical skill, but project lead as well. I like his energy tho, i would be curious what it will be like to work with someone like Ved.

[@xiwenliu1572]

Should have watched this before today's interview.

[@griesrt]

Ved, if we want to calculate the related jobs, not based on the latent factors, but the content of the job what types of content embeddings would you use?

[@gemamiable]

You mean like using content based recommendation? Assuming that the content of the jobs are like the role and the requitements. How about starting from BagofWords or TF-IDF ? Or maybe you want to try more advandce word embeddings like CBoW or skipgram

[@griesrt]

@@gemamiable Yes, those are good options.

[@prashantgupta1734]

Very well said at 39:08

[@malice112]

Just to clarify when Ved refers to a User feature matrix would that be collaborative based filtering and when he refers to an Item feature matrix would that be content based filtering?

[@leisana4097]

Question asked was given a description of a job show top N jobs. But the candidate went along with User and Jobs relations. From where did the candidate arrive at the User features. This problem is related to given a document get the top N sentences or more specifically given a description of a job generate top K jobs. I didn't quite get from where did the features of users come into the picture.

[@TheXplorian]

The point is after getting the matrix of user-items interactions, in the process you'd have learnt embedding matrices both for users and items, after that you'll be able to use the items matrix/Embeddings to apply k nearest neighbor using the embedding itself, the point that Ved missed i believe is the fact that we won't have Embeddings for new items, as they won't have previous interactions, as they are new, so they won't have an embedding vector to be compared against.

[@leisana4097]

Let's ask another similar question : Given a description of a product generate images related to that product description. Say a description like , a red round fruit , the related answer would be image of an apple. Another one could be a 'Vehicle with four wheels' now a car is more relevant than a bike. you can generate BMW, Ferrari, Ford etc certainly you won't generate Harley Davidson bikes. You don't need user preferences or features like male , female , age , what product he bought etc. Why do you want to bring in the user preferences, the user could be someone who is a doctor , a banker which has got nothing to do with whether he likes apple or he prefers cars to jeep. It's very much related to the description. That's what I was trying to tell.

[@jayaramkollipara7257]

Hey Jay & Ved, could you shed light how do you actually initialize the User-Job Rating Matrix to start the Colloaborative filtering algorithm ?

[@manishsharma2211]

Brilliant

[@Acha413]

I have one query, if there is a new job and we do not have a vector representation for it, this is similar to a cold start problem, how do we recommend this job and to which user.? one of the very simple ways I can think is to convert it using a doc2vec method to give it a vector representation, use cosine similarity (in doc2vec space for other documents) to find nearest jobs, then consider the nearest jobs and use their features obtained from matrix factorization to compute the new vector for the job, maybe it might be useful, anyone tried...

[@paragjain2762]

Hi Ved and Jay, thanks for doing this. Do you mind throwing some light on ways to calculate the user/job matrix ? Also, how would you go about creating the f-matrix at 13:00

[@shairuno]

F matrix is a dense matrix. Typically, people use either gradient based methods or alternate Least squeeze to figure out the entry of the f matrix.

[@weikangduan9852]

ved so handsome!!

[@lalkakonus]

Booring

[@be051268]

what a trash video for interview and teaching

[@janet-matsen]

As a woman, it is hard to listen to the interviewee describing the job seekers as "he". Come on, Bros!

[@hihi-wc7dx]

She she she happy now?