transcribe

How He Aced Meta Data Science Interview πŸš€

DataInterview Β· 45m Β· transcribed May 2026
More from DataInterview Business
𝕏 Share β–Ά YouTube πŸ“₯ PDF πŸ€– .md

Transcript

0:00 hello matt hey dan how's it going i'm good how are you hey thanks for partaking in this uh in this data talk as a guest um matt here is a um upcoming data scientist at meta um he used data tv.com um especially the uh online subscription course as a way to prepare for meta's data science interview and i basically invited him as a guest to basically give him a chance to talk about his career experience

0:30 what are some cool things that he worked on and then basically talk about some general interview prep tip for data science interview and along with meta's interview and then as a bonus content um at the end of this video what we're going to do is we're going to go through a case problem together um based on what's that so definitely watch this video end to end and and i hope you find this video helpful so without further ado matt i'd

0:56 like to give the floor over to you and could you just talk about your overall career journey for those who do not know who you are yes sounds good and uh before i get started i just want to say thank you for putting together the data interview platform um like as you mentioned it was kind of critical and i think my success in the interview but a little bit about myself so i've been in data analytics

1:18 probably around five years at this point started off in consulting got a kind of a range of experiences and you know building dashboards etl development some machine learning type stuff uh the the hours kind of got got to me a bit so recently i joined a fintech company so in the finance space dealing a lot with like stocks um in my role i build a lot of like dashboards work with a lot of stakeholders across the company to

1:47 solve problems with data and recently i've had the urge to get into a true data science career and uh really with like a product focus and that's what really drew me to uh to meta and some of the products that they have in their portfolio so uh i gave it a shot updated my resume and uh went through the interviews and stuff and you know things things panned out in my favor and i'm happy to be joining

2:11 them in a couple weeks here that's that's awesome yeah um it's always very interesting to see people with a similar kind of route that i've also personally taken because um i looked at your linkedin pro profile and and uh for those who are interested in definitely connecting with matt um i'll provide a link to his profile down below in the description of this youtube video um and i one thing i that really caught my attention was that um you know

2:38 personally for me i started out as a as a consultant as a at louisiana hamilton and eventually i got into the fintech company at paypal and then i got into the uh really you know one of those fan company which was google for my case so i saw that you know you started your career in kpmg and then you started over to vanguard and then now you're going to be joining meta pretty soon so it was

3:00 very interesting to see that similar parallel there um so i'd like to kind of start with maybe like the early days of your data science career like how did you kind of start your way into data science anyways um i think this is really helpful for you know aspiring data scientists who want to actually break into the field definitely yeah so i did i did have like a data science i guess background in undergrad so i went to a

3:23 college called saint joe's outside of philadelphia and they actually had a program called uh business intelligence and analytics and it was kind of like your your intro to data science course if you will we got we touched on everything from like intro to statistics to some data mining concepts to some intro to like sql and database type stuff so you kind of got that well-rounded backgrounds and after kind of going through some of my courses and

3:49 in my major i was like i love this stuff definitely want to continue on with my career where can i get a job in it so that was kind of what ultimately led me to that and i think one of the big things i was looking for out of college was as much experience as i could get as quickly as possible um and that being said i think consulting fits perfectly into that kind of goals like hey you can work with you

4:13 know 15 clients in your first year of work all in different industries on all different sorts of projects so i think as soon as i got a role at kpmg i was like this is awesome it kind of aligns with what i wanted for my my first job right this wealth of experience and then later in my career i felt like if i had all that broad experience i could really know quicker what i want to do and which

4:37 pillar i really want to like specialize in or you know pursue more full-time and that kind of that's kind of been like the driving factor i think in my career so far is like get as much experience as i can and then kind of like filter through it and write what do i where do my passions lie and then like pursue that next so that's been kind of the path i'm on and uh it's taken me to to meta in a few

4:58 weeks so really excited yeah that's awesome um so what were what was one really cool thing that you worked out at pmg yeah so one of the the cool things we get to work on um this was more like an internal project but i actually worked with some senior like partners on developing a recruiting uh like analytics and dashboard suite where these partners as as they're going to campuses across the country to do like recruiting type

5:28 efforts um get new people to the firm get people interested in the brand um i actually built out this like dashboard tool and kind of like pipeline of of gathering data from all these prospective um students and across the country and we were able to like plot it and almost have this real time like heat map that partners could show uh you know as they go to your campus they could select like a drop down it

5:55 shows you alumni currently at the company it pulls up their profile information some analytics around like tenure and uh the industries they support so i think one of the outcomes of that is like students immediately had contacts that shared a similar background that they could kind of reach out to connect with on linkedin to go to networking events etc so i think it was a really cool experience because it kind of connected students entering the workforce with alumni um

6:23 already in the workforce so it's kind of a really cool experience yeah that's awesome yeah that's awesome that you know you're providing this type of experience where you can bridge the basically this connection between the um the students who are trying to get their first entry-level job and they um the recruiters and the hiring managers um yeah so um i'd like to kind of talk a little bit more about like maybe your bank experience at vanguard so i saw that um

6:48 you know you built some data visualization over our vanguard it seems like there were some pretty cool like analytic based projects that you were working over there so what did you do over there yeah definitely so right now i'm on like a data science team we support a lot of different teams across the enterprise as you can guess like being vanguard we're a finance focused company so a lot of our use cases are dealing with like

7:11 finance financial data um you know clients account information monitoring for like fraud detection etc one of the cool things that we as a company are interested in and i think the industry overall is the idea of process mining so how can we leverage this log data generated from you know dozens of systems internally externally that we're collecting data on how can we like make sense of this and start to identify trends in this log data so one

7:41 of the things i've been working on for the past year or so is the tools around process mining basically sifting through millions and millions of log type data and starting to get that full lineage of what happened what's business as usual look like what are variations of that can we detect these anomalies maybe in real time before they kind of lead to better things got it got it that's been really interesting by the way um back when i

8:09 used to work at google the foreign contributions but this is actually managed by vanguard oh really yeah yeah yeah exactly exactly and now as a self-important individual you know i need a i'm trying to establish my our individual department account along with us uh basically self-employment um uh 401k uh contribution account so i'm planning to use vanguard actually definitely yeah it's a great company yeah it is um okay so that's that's really interesting um

8:41 so so now i'd like to kind of segue over to just a general interview prep um that you did um so how did you kind of start your process of i guess uh jump starting your uh interview prep for these companies yeah it sounds good um so i'd say if i had to like time box my interview prep i'd say it was around two months of like full force dedicated kind of prep time and the first week of that was a lot of like

9:07 what i call orientation type stuff basically going out to linkedin searching 15 to 20 different job descriptions for data science and starting to see in my head like what are the commonalities between these roles where i can invest my most time studying for the greatest return so you know every role had python every role had sql um some roles had a b testing some didn't um so i was really just trying to get a sense like what the

9:33 field's looking for in terms of skill set um to help me guide my my studies i'd say another thing that i did in week one was start the mock interviews early uh i can't stress that enough i probably did over 50 mock interviews at this point um just to kind of get up to speed on that and the first few were absolutely horrible you know i was fumbling over my words i didn't have any structure i was

9:57 i was rambling so i'd say like the biggest thing i could stress would be start the mock interviews week one and going back to your platform like that slack channel i met so many individuals who were interviewing for similar uh product data science roles at meta specifically um that i was able to mock interview with and you know a few of them also got the role so i have some co-workers to to meet in a few weeks but

10:24 i would say do the mock interviews as early as possible for three reasons one is you obviously get better at learning the structures learning how to talk through these ideas in real time on the spot which is critical i'd say two is you get a gauge of your competition right i i mentioned i was terrible when i first started after a few weeks you start to see yourself like grow in terms of your responses your your answer structure etc and you also

10:52 get to meet with some really smart data scientists already who are at this level so it kind of is a really good gauge of where you're at where you need to get to and how you're progressing um over the weeks um and then i'd say third is just to me like you know expand your network was i met a ton of people who ultimately got jobs all over some of them at meta but uh i've kept in touch with them for the

11:15 the past few weeks too so it's it's been yeah yeah no it's um it's been very interesting to see how you know people would pair up uh you know with for these mock interviews the slack community group that's managed um you know the data interview service and um you know uh yeah like one thing that i definitely encourage for you know any any candidates who are currently prepared for data science science interviews is that you know more often than not they

11:41 they'll kind of think about it as like okay i'm going to read all the resources out there i'm going to practice on my own and then maybe i will do some walk into you and then do the um interview you know basically the mentality is like okay i'm going to be 100 prepared and then do the actual mock interview or the actual um interview um don't do that you're only building up unnecessary pressure for yourself um rather do those

12:05 in sync in synchrony of one another so do this in peril because what it's going to do is it's you know you're getting a feedback loop you know uh you're starting with the minimal product which is basically you know yourself not clearly knowing what you're going to expect in these market review setting or the actual interview setting um but then once you go through it you know not only you get that anxiety and the fear of someone out of the way but

12:31 at the same time you're learning oh okay these are the topics i definitely need to hone in on so this constant feedback loop is going to be ultimately helpful in terms of the in terms of the um being being very effective yeah 100 i'd say one thing to add on there is like after doing so many mock interviews you get into the actual interview on site and you know you're just relaxed right you've done this so many times before

12:55 where your nerves traditionally in the past like my voice is shaky like you get anxiety um there's like no better feeling in the interview when it counts like being relaxed calm and be able to think through and yeah the only way to get there is through mock interviews yeah and the other very important thing is that you know yeah like you're in your zen but also at the same time you're bringing um you're almost sort of making the

13:20 interview not like okay there's the interviewer an indie sort of thing but it's it becomes conversational you know um and and that's one thing i just want to emphasize so much because i recognize that you know either whether i'm working through clients or um you know whether i watch other content out there sometimes they really emphasize this like um approach the interview question with this framework you know here's step one this is what you need to say step two what you need

13:47 to say uh yes that kind of structural response matters but what matters a lot more is whether you can make it conversational you know are you kind of engaging it in a way where you know if you are an actual data scientist at that company and this interviewers was your 10th year was your colleague you know how do you interactively kind of like work through this problem together that's what matters more you know yeah um yeah so that's really cool um so are

14:16 there any kind of general um you know technical crap like sql or coding sort of things that you that that um any kind of advice on that in terms of how people can prepare for those topics yeah definitely that's a good point so i think for the majority of the the interviews i went on meta and others the first round is usually that screening route right where it's a technical with the case study and everyone knows there's a bunch of other

14:44 like stat and probability stuff that is also part of these roles but i would say front load your study with sql and case studies and the reason is like if you don't get past the 10 or the tech screen you obviously don't get to the onsite round so it's like you got to be really good at these first two things before you can even focus on the other ones um and one of the ways i did that was i did probably at least 30

15:07 to 60 minutes of sql exercises a day just to like stay sharp you want to be able to write out window functions like yeah in your sleep you want to be able to you know differentiate between union and union all like on the spot and recognize those patterns and you know the only there's no secret to it the only way to get there is uh investing the time and every day consistently yeah yeah you know that's a similar um tactic

15:35 that i use um to prepare for uh interviews that were that had these sql and coding problems as well uh so personally for me um you know kind of like you you know basically i invested about 30 to 60 minutes on a daily basis and i'm solving about like four to five problems a day yeah and i did that like starting the first month of my prep you know um there's really no magical bullet on this

15:59 one it's just it's just a matter of practice again and again you know yeah and i on that point i've seen some candidates who i've mock interviewed with um you know there's there's limited prep resources out here so if you do enough of them you start to memorize the answers and i would highly suggest like don't try to memorize them because i've seen candidates you know recognize i've seen this pattern i've seen this problem before how can i solve it and

16:23 you know the problems you're going to get on the the interview aren't any ones you've seen before so if you are strictly trying to memorize things um i would suggest don't do that try to understand the concepts exactly yeah yeah focus on the concept you know build out logic um and then from there you're going to start recognizing patterns of how to crack these uh sql uh or coding questions yeah definitely yeah um okay and in

16:50 terms of i guess the general like you know obviously uh one thing meta loves to cover is like product sense questions but also like other uh product data science roles of different companies as well you know they love to ask these product sense questions how are you going about preparing for these products those questions yeah definitely so i think in general when you're talking about like product sense case studies there's like three or four buckets and

17:14 really comes down to like creating a metric right being able to understand all the different parts of a metric and what you're tracking so that's like one of the type of case studies is around come up with a metric to measure success measure value measure something the second type of case study is like investigative right so the business has key metric and it's declined it's increased it's spiked find out why and the third one is around designing an

17:41 experiment so being a product focused company a lot of them want to test how what are your skills around designing an experiment to you know drive a change in a metric for the business so those are like the three types of core case studies um i've seen some companies ask one's around like how would you improve a product those those are also really interesting but i feel like there may be more on the product management side of things from

18:07 what i've seen um for the data science roles it's really this first three yep got it okay that's cool um all right um do you have any final sort of general tips for candidates out there in terms of how they should prepare for data science interview before we segue to the next part which is actually going through a case problem together definitely yes i just want to reiterate start your mock interviews early regardless of you know

18:31 where you think you're at the first one's going to be probably a little rough but you'll get so much better every day and it'll help you tremendously in every round of the interview um secondly practice sql every day i know it's uh sometimes not the most fun thing to do but that repetition and consistency will you know pay up pay dividends in the tech screen when it counts that's awesome all right thank you so much for for that tip um okay so what

18:58 i'm gonna do next is i'm gonna share a screen um i also actually give you a link to this uh this google uh google document and what we're gonna do is we're gonna flesh out this document together okay all right okay so this problem is basically um involving the whatsapp um and uh the problem is you know how would you detect spammers on whatsapp

19:31 definitely yes this is an interesting problem um i think one of the things i'd look i first try to do here is start to ask clarifying questions around what's a spammer so the way i'm kind of thinking about spammers are users who are not on the platform to engage in conversation right they're trying to push external links they're trying to you know sending massive group chats to as many numbers as they can get um and spamming

19:59 you right that's what's being done so one of the first things i would do is is try to really narrow down on what's spam mean in the context of this problem yep yeah that's a that's a really good point um and i would say that you know i do agree with the uh with the idea that um you know basically there's a bit of this like malicious intent behind what they're trying to do and so what

20:21 spammers will typically do is that it you know in a form of like a automated fashion or um they're almost sort of copy pasting the same message again and again to as many users as you can um and typically the message content itself will have some sort of link and and it's like a clickbait essentially you know it might say something lines up um you know go to this link because it's going to give you 30 or

20:48 something like that right and once the um once the victim goes through that link you know they might enter some like credit card information or something like that next you know they get their um you know uh important stuff like stolen right so so i would say like that would be like the nature of what spammer is definitely yeah that makes sense um the next thing i try to do is come up with some features right so how we know what

21:14 spammers are how are we going to think about tracking this and you know knowing some data points we can look at to give us confidence one way or the other um and i kind of like to think of features in terms of like buckets so the way i'm thinking is there's three buckets in my head one is metadata so you know data around that phone number or that user account um and i can flesh out some additional

21:39 features in a minute the second one is more transactional or activity so what have these users done what evidence do we have you know based on activity and then the third one is more of like an aggregate ecosystem metric so where does this activity compare in terms of all of our users so those are kind of the three buckets i'm thinking of in in thinking of features and how you might be able to leverage this data to create these

22:07 features and identify spammers got it one real quick what was the last point that you just mentioned there um ecosystem like aggregate level uh metrics okay got it yeah aggregations yeah got it okay yeah that makes sense um yeah so i think one thing that i think these are really good data sources um you know there's definitely a lot of information that can be useful um [Music] i also think um one thing that we need to consider as

22:39 well is also how we would actually get the labels in this case right yes um because in order to uh build out this you know basically what this question is asking is you know whether you're building a rule based approach or machine learning approach how would you ultimately apply a user as spam based on their behaviors right um so in order to do that we need to think about how we're going to actually get the labels

23:06 yeah that's a good point so the labeling the labeling question i think can be answered in two ways one is we can kind of have this like human in the middle approach where we present them data for perhaps manual review and they label it saying nope this looks good or yep this is spam let's let's label it as spam the other way we can potentially use it at is on the initial past use some rule based approach or perhaps get our hands

23:35 on existing data that might have been manually reviewed by some security team um to identify you know patterns that historically represented spam on this on the whatsapp platform yeah and i think another really good approach to getting labels is actually crowdsourcing it from the users themselves um because what happens on whatsapp is that when you send out a message um uh when you receive a message and this person isn't in your contact list

24:07 then a pop-up is going to appear in the message saying um you know his system is like or do you accept this contact uh or is do you want to report this as a spam uh so so that's another critical information that we can get um by having users actually generate those labels for us definitely yeah that's a that's a great point to include um historically sometimes i am hesitant for user-based uh feedback but i think in the case like fraud and

24:37 stuff as you mentioned like users are more than happy to stop people from spamming them so they're usually more likely in this survey case than you know other survey type cases on the on platform yeah yeah uh but it but also if we think about it um you know in the in the world of like um credit card um like fraud detection because that because i came from that world before working at google um is that um you know you get the labels

25:05 actually from those from the victims you know um because the victim was report saying that hey you know i got my credit card um information store and so that's how you would actually um accumulate the labels over time um and then eventually that's once you have some label initial set of labels from um you know that are generated from the users um there's a essentially manual review reviewers who are going to generate some rule-based approach and before a possible credit card fraud

25:37 can happen they will look at a user's profile along with their transaction history and then pre and then preemptively you know block this user um based on whether they're they have they're likely to be fraudulent or not uh so i think i think a combination of all the three things that we just mentioned are ultimately one way uh our ultimate you know ways we can get labels yeah definitely yeah um okay all right so

26:08 i think we um have some general kind of prerequisite you know the information that we need in order to start kind of um solving this problem um so so now we need to think about the methodology of how we're going to actually detect spammers yeah so there's there's two ways i'm kind of thinking about being able to detect this one is given that we have the labels we can leverage like a machine learning technique uh specifically a classification type model

26:37 thinking like uh logistic regression here maybe a random farce or decision tree model if we don't have any labels and say we're this is the first time as a business we're looking to solve this problem we might be able to leverage some basic statistics and distributions to identify populations of potentially spammers um so do you have a preference on where i start so let's actually assume that um so we essentially gotten the labels um for the the three

27:09 things that we just mentioned there and so um i think the next thing that we should definitely cover together is uh how we're gonna actually build a classification model um and one thing i just wanna provide there is that you know obviously um this the span in this case is going to be a low probability event model right so um so we have to think about you know how we're going to consider the fact that we're dealing with low

27:34 um true positiveness in this case um you know compared to um uh messages that are not spent yeah definitely so anytime you're dealing with like fraud as you mentioned you have the the inbounds class issue um so that should that raises a couple concerns in my head of in terms of how we're going to evaluate this and what are the metrics we we really want to look at but going back to that going back to this firstly is like how

27:59 are we going to deal with this imbalance from building a model perspective there's a number of techniques we can use to help address the bias that's introduced by this imbalance in in class labels one that comes to mind is like it's called smoke um it's like synthetic minority over sampling and what we're doing here basically is over sampling our our true positives right our smaller class here to try to make them more prevalent in our training data set so

28:28 our model breaks them up and we're able to develop something meaningful to capture us in real world yeah yeah yeah and then there's also um the down sampling um as well which is basically removing some of the non um spammer used group in this case as a way to bring that ratio a little bit more optimal for building a model yeah that's a good point

28:58 um so that yeah so that's definitely something that we should consider as well and i think uh we also briefly discussed what type of model we're going to use in this case and i think um you know religious regression and random four should be fine um yeah go ahead just jumping in here uh legit regression random far obviously both classification models um traditionally random farce is usually better performer and the reason why it's known as an ensemble model so actually

29:29 under the hood it creates a bunch of individual decision trees and it is able to kind of like aggregate all these decision trees predictions and roll it up into a single one so most of the time a just a random farce is going to be slightly better than logistic regression but yup yup definitely a good point the point um okay so in terms of how we're going to actually measure the performance of the model um obviously in this case

29:58 um because we're dealing with imbalance you know we don't want to use accuracy we want to use metrics like auc in this case um another set of metrics that we should definitely consider are um f1 score precision recall definitely yeah i totally agree with you there yeah um okay all right so one thing that we should also consider is um at what stage do we basically flag

30:29 a um a spammer uh a user as a spammer right um so if we think about it there's a user journey around this right so basically a user is going to sign up on whatsapp and then they'll start initiating messages right so we want to basically use either something about their profile information during the sign up stage or during the time when they're actually messaging and based on the way they message

31:02 basically flag this user as uh as a spammer so ultimately we need to think about what the prediction point is yeah um do you have any do you have some initial thoughts about that yeah so that that's the interesting point um i can kind of think of it from like two points in the user journey one is if this spammer has been seen before as a spammer we can obviously stop them before they are even ever able to send a

31:28 message right yeah so there's there's that point um and then two is almost like it's the first or the prediction point i think would be sending messages over a certain threshold and the thinking here is like there's a ton of normal you know users who are sending messages but once they send too many messages in too short of a time there's that fine line between regular user and probably spammer so we can have this almost uh

32:00 like threshold at the sending message stage uh further prediction point yeah that's a good point um yeah so in fact um generally speaking you know when we think about like building a broad model um it's not that the fraud model will there you know it's not that a single fraud model will only tackle the entire lifecycle of a fraudulent user um there's a fragile there's a broad model that's built in for every stage of a

32:32 user journey so we're talking about the signup stage and then maybe the first event stage the third event stage the fifth event stage in this case right so i like i definitely like the idea of first of all targeting users at the res red regis station stage so so basically block them out even before they will send out a first message right right and um key information like um have they been blacklisted before uh well how are we going about doing that

33:02 well um if they've been blacklisted before then we would know what email address the phone number ikea addresses are associated with that blacklisted account right um so we can use those type of information as a way to um basically flag a user flag a new user as as a spammer in that register registration stage definitely yeah yeah and i think um another key information we could definitely use in our

33:32 aggregations as well so we can look at for this given user's ikea address how many accounts how many new accounts did they try to generate the past you know one day or seven days or 30 days right exactly yeah exactly um and and you know compared to normal user might only establish maybe one account for a given ip address spammer on the other hand their intention is to create as many accounts as they can so they can really

34:00 disseminate this clickbait messages as many as they can so um they'll have a very high um you know uh values in terms of that type of aggregation-based future right yeah um and then yeah so i think we could definitely build a model at the registers stage and then the second model we could potentially build is basically the early activity states like what you mentioned which is um let's track their behavior um in the first couple messages and see

34:32 whether are they generating a lot of quant a lot of messages over a short period of time yeah i agree and just to add on to that early activity stage going back to like your basic statistics looking at a distribution of messages sent by user per day um i think we'll give you a really idea of what's the norm on the platform to begin with and then from a a spammer percentage maybe you select like the 95th or 99th uh

34:58 percentile as like that threshold that you know is really good evidence of of spammy behavior i think uh you know going back to simple statistics would would help a lot in this case yup definitely um and i think other really good information we could consider as um you know in a way like it span like is especially those like very clever spammers what they'll do is like they'll use um automation right like automate messages and so if we

35:29 think about what is an organic behavior on you know messaging service is that a user is going to type something and and oftentimes you know when you're in whatsapp or messenger you'll see this like blog indicating that someone is typing something right um but a spammer on the other hand though either if they're doing it manually they'll just copy paste things you know or don't use automation with a way to just send a message very quickly

35:54 right so so perhaps the average time it took to send to actually type up the message that that could be another key information as well yeah i think that's a great point like the the tts would be almost zero right if it's just copy and pasting bulk messaging so i think that would be a great metric yeah yeah um and of course i mean you know there's there's a there's dozens of other potential signals that you know we

36:21 could we could you know generate um but but then that says you know i think this kind of captures the heart of i guess you know what spammers would do and how we would actually going about um you know detecting protecting this family in this case definitely yeah not just one one more uh data point we could look at just for the the viewers is like messages sent versus messages received idea here is spammers don't care about

36:45 responding to message right they're they're going to be so skewed on the sending front um that they will probably have close to zero messages received um so that would be another i think dead giveaway of a spammer yeah yeah that's uh that's a that's another really good signal as well um okay yeah so so obviously you know there's a lot more that we can um cover through this uh topic but in essence i think this pretty much captures the idea

37:12 of how do you go about approaching a problem like this um thank you for your participation in this one cool yeah this is a fun one yeah um okay so before we um kind of end the sort of practice session together um there i do want to maybe um kind of go through some behavior questions um together you know obviously people people do have these behavioral rounds in these interviews and i think it's a really good kind of demonstration so um so we

37:45 can kind of take turns um is there a behavior question you'd like to ask me and then i can also ask you as well yeah definitely and i'd say just to like preface this one of the things i think they look for in a lot of these interviews is data scientists are kind of these like rare employees who are super technical super business savvy and can operate kind of in between and they really look for that

38:08 like cross-functional um type answers right of you engaging with stakeholders as well as like demonstrating uh technical knowledge so i'd say like as you're thinking about some of these behavioral ones think of um experiences where you know they involve multiple teams multiple stakeholders multiple technologies and try to like answer the questions in that lens yeah that's a really good point i'm going on to the question uh one that i like to ask is like what's a recent

38:36 project used to influence with data yeah thank you so much for asking that question so i would say the most recent project would really be the last project i delivered at google and so i work with the financial analyst over in the technical infrastructure team especially within the network pillar so the technical infrastructure at google they'll basically they basically deal with you know building out data centers to uh support um you know basically internet traffic for

39:05 like google search uh cloud youtube uh google maps and so forth um so it's a very critical functioning team and the financial analysts over there you know what they really what they're really concerned about um expenses and usually expenses can break down into two factors in operating expense or capital expenditure for cap x and what capex it basically is is that you're gonna invest maybe one million dollars um tomorrow because you expect the network traffic to be

39:37 much higher you know uh because if you think about youtube and and um google search you know they're on a very sort of expendable growth in terms of the track that is required to support that right right so it's like an investment for tomorrow and what they're trying to calculate is how much should we invest this year the next year uh basically a fiber horizon and so traditionally what the financial analyst would do is they would have like

40:04 a manual approach like some business heuristics like very clever logic and excel spreadsheet as a way to estimate this right and as i was a fine as data scientist google one thing that we need what we propose is like we can definitely automate this for you guys and we can build a much better model um you know through machine learning um and so this is really an opportunity for me to kind of lead that sort of okay

40:29 this is how much we should expect and you know hopefully it's also helpful in terms of your planning process as well so i um and started by getting all the data sources that i needed so i had the historical capital um capex um spending um for the to be able to support this network traffic um and then i use this time series model there was a lot of models that i was kind of tinkering with it took a lot of trials and errors but

40:59 all in all what i ended up using was model called bayesian dynamic linear model basically it's just a bayesian based uh linear model and it allowed this trajectory of you know what this five-year horizon is and i use time series cross-validation as a way to show that this model is actually better than the business based line um and so what once we have proven out this capability we rolled it out um and the financial analysts agreed to use this for their um

41:31 2022 uh planning so that wasn't really uh that was a really good um kind of experience of being able to deliver something and also just kind of lead that part of that process that's awesome yeah so it sounds like they you know you presented an idea and they actually ended up implementing it for the next year's predictions yeah yeah that's awesome yup okay um actually so i'll ask the same question over to you so could you give an example where you let

41:59 something with data yeah definitely so um i mentioned earlier like there's a big topic in the industry known as like process mining which is basically gathering insights from log-based data from internal systems external systems etc so one of the projects i was working on was trying to come off the prototype to showcase like the the strengths of process mining as a tool and a methodology as well as kind of getting adoption from these business teams on

42:29 how you can use the output of this to you know drive a decision or you know improve a process process etc so as part of this project i went around to various teams in business started to talk with them about what are some of your gold source systems that you're leveraging in your business generating data from using as like workflow management type tools start to get a list of them get some of the technical details understand the

42:56 data etc and put together this like prototype of a dashboard where you can see in real time all the transactions that have went through your system what the path they followed what's the most common path and what are some of these like variations or potential anomalies look like so at a quick glance you could instantly see kind of everything going on in your business area which i think was like an unprecedented lens that they had and going on to like the adoption piece

43:27 i needed a way to sell this back to teens right so the so what question so one of the things i tried to come up with is a metric that was easily communic like easily to communicate to these teams um and something that was a pain point to a lot of teams which is manual review right no one likes to do manual review it requires hours of time it's outside of the automation process by definition and everyone wants to reduce it so i

43:55 came up with this metric called mer which is manual review rate and i use this as like a proxy to communicate the value of process money to whatever business area i was talking in so that was like kind of my my generic metric that i go into your business understand what manual means in your in your case and try to like fit that lens through there to say hey if you look at this data we can actually reduce your mer by

44:24 x percent by reducing these processes so you can see what is causing the the uptick in that metric and then you could also see the specific steps that you could potentially eliminate or you know refactor due process improvement etc um and that's been like a huge hit with the businesses is getting that lens understanding what to do with it and then taking action that's really cool that's uh yeah thank you so much for sharing that um okay so

44:54 uh there you have it guys um you know uh we definitely covered very awesome kind of topics uh from the beginning to the end of the skull um okay so uh so you know thank you matt so much you know for partaking in this um and for those who are watching this um you know if you're looking for data science interview prep uh check out data tv.com it contains courses and coaching services that can be helpful for your

45:22 interview prep and you can also email me at dan at data dot com if you have any questions about that

Summary

Matt, an upcoming data scientist at Meta, shares his career journey and interview preparation strategies, emphasizing the importance of experience and mock interviews. He discusses his transition from consulting to fintech and now to Meta, highlighting the skills and projects that shaped his path.

- Matt utilized data.tv.com to prepare for Meta's data science interviews, emphasizing its role in his success.
- He started his career in data analytics and consulting, gaining diverse experiences before moving to a fintech company.
- Key projects included developing a recruiting analytics dashboard at KPMG and working on process mining at Vanguard.
- For interview prep, he recommends starting mock interviews early to build confidence and improve skills.
- He emphasizes the importance of SQL practice and understanding core data science concepts rather than memorizing answers.
- Matt outlines common types of product sense case studies, including metric creation, investigative analysis, and experiment design.
- He discusses strategies for dealing with class imbalance in spam detection models, such as SMOTE and downsampling.
- The conversation concludes with a collaborative case problem on detecting spammers on WhatsApp, discussing features, labeling, and model evaluation metrics.
Β© transcribe Β· For agents Built with care and craft by Gokul Rajaram