Text By the Bay 2015: Nima Sarshar, Using Big Data to Identify the World's Top Experts
Recording: Text By the Bay 2015: Nima Sarshar, Using Big Data to Identify the World's Top Experts
well welcome hello uh thanks for stopping by my name is Nima SAR I'm the CT of uh empowered which is not too far away right across AT&T Park um so I'm going to quickly go over uh what we do in the tech we've built to identify World experts in in any topic basically and The Business that we've built around it and the challenges we have and uh you know hopefully I'll go quick through that and then we could chat a little bit after so uh I mean I probably don't have to argue too much for this I mean there's a lot of crap out there if you go to the internet I mean it's just really hard to get real use for information actually is pretty interesting skill that not everybody has and it's uh hurting consumers it's hurting also the advertisers so I was just like like looking at something for um American Airlines and I go to this page and I see all this junk and uh but all I wanted to do is was basically I was just doing a little bit of a research about American Airlines right so I have to now make Oprah cry which how can you right and then there's an ad about a pretty good actually uh snowblower uh and there's all this crap at the bottom that that I don't know what it is but uh so that's clearly not helping anybody uh so the consumers cannot get the information they're looking for if I'm researching about getting the best airline or what's a good Airline uh you know miles for which airline to pick this is not helping anybody um also those advertisers I don't see how that snowblower as good as it is is benefiting from being there so that's what we're trying to solve so consumer is not finding the information they want easily um they feel frustrated sort of dirty after coming out of this page uh then uh and obviously the brands are not getting um getting what they're looking for so what a way to to go about that is what we've started and empowered is why don't I clean up all that mess and if you're looking for a certain topic why don't I have a clean way for you to go and get the information you want about that topic so we already do that in a way we have all these Brands let's say CET that compares this versus that a lot of us go there to research if you want to buy a gadget sort of somehow we trust that brand but what we've tried to do is that the trust should be with the expert I mean if you're an expert regardless of where you write if you're writing for a small block I want you to give you that credibility just because you're writing that it's not coming from the brand itself so previously the brands hold that they filter people we assume we just trust them but why don't we kind of do that in a more organic way so to me this could be a scate for everything but machine based and and independent or you could think of it as a Reddit for not non nerds right so we kind of know who are experts in different topics um un ready we know how you know people just bubble up because we think they're experts because people validate them what if we did the same thing for everything um and so this is sort of a dashboard that we have I mean this is a little bit cheesy it's Apple watch but you go there and you get the story of the day but on the left you see people that are experts in fields related to Apple watch it could be iOS devices Smartwatch wearable technology and so forth things that are categories that are related to what you're looking for and then you could now what you could see is now we're not yet doing that for the love of God we're I'll have a business out of that so what we do is basically we also allow the uh brand to to see what experts have to say about the products that they've launched and if they're things that they like we actually turn that into a native ad like a promoted story that you would see in your Facebook so uh here's basically the core of our business uh and the left B this is an actual campaign we ran for Motorola when they came up with Moto X um so somebody who is an expert in Android smartphones thinks that Motorola's new Moto X might be the best Android iPhone I've seen all year that's an that's a you know that's an independent story she's been ranked independently because of all the things she's written in the past and she has something pretty good to say about Moto X so we turn that into a a promoted Story by empowered when people click on it the ad is the story so you just go there and you just read for yourself the value we bring to the consumer is that Badge of trust and again that's been independent nobody's paid her to write the story and this is based on everything that she's is written a long over a long period of time so that becomes a very powerful ad we believe the consumer benefits because it's it's a actual piece of story it's an and nonetheless I'm I'm not delusional but also clearly the the Brand's benefit because now the trust that they they could not gain through their own branded ads they could gain by independently organically generated what we call earned media so but who's an expert uh so like you could be philosophical about it but we've been pretty principal we wanted something practical that we believe in so to me an expert is somebody who writes about a topic consistently and in depth so you have to write a lot uh you have to kind of it needs to be your job but you can also be writing about everything if you're if if you're an expert about everything you're expert about nothing so about a small set of related topics we want you to be writing have a track record of writing about that that but also no matter how much you write if nobody cares then that's not interesting either so we want people to engage with your content validate it somehow share it comment on it and so forth and the third piece which I think is interesting that we started putting in is opinions so you could break the news and get a lot of be the first one who breaks the news and get a lot of tweets and shares but as a consumer it's not benefiting me that much right so what that you're the first one who who found a leaked photo of something um some product I want you to actually Express opinion I want you to put things in context for me so we want so that's basically the third piece that goes into how we rank people um and at the end of the day uh and I'm as I showed you there I claimed the lady was number three Android smart or Android smartphone expert and so forth all these rankings are there but how do I even measure that right um so what we have what I call is um the business of Engagement I want at to me at the end of the day if I put that story in front of you whether you read it or not is the greatest measure for how good that article was so so you were looking for an article I thought this is going to benefit you you clicked on it you went there and you bounc back that is not good so we've kind of put our money where our mouth is so we Define we only charge the brands if somebody reads an article at least for 15 seconds Okay so then it's in my it's sort of in my interest uh even in the business side is to come up with with a good way of gauging whether this article is engaging enough okay so that's why at the end of the day our rankings are correlated with the probability that somebody someone and random would actually read that article at least for 15 seconds and what we've came up with as you see here is basically if somebody is ranked one in a given topic the three and a half times more likely it's three and a half times more likely for their story to be read at least 15 second versus somebody else who just wrote something about that article occasionally so to me that's sort of the greatest validation but the beauty of it is that that's exactly correlated with our what our businesses so now let's get to the a little bit of little bit deeper um so this is what I mean by consistency in depth this is um this is a 100 tweets of a lady at some point in time but you know it's hard to I mean she's a celebrity but you can kind of figure out what she's into so to me expertise versus influence really the core of it is is is depth in a you know where do you have depth in a given topic now she's pretty famous anything she tweets gets gets engaged with but really looking at those the cloud word I mean you don't know whe but but this you guys probably recognize it's it's somebody who's writing about Big Data data science something like something along that nature um and that and she's like our I don't know I mean you you may argue with the with the ranking or not but but she's clearly somebody who consistently writes about data science subjects now to you guys the big data in spark and Cloud era and Hadoop and all they're all related to each other but we need a sort of a topical model that understands that so the first order business is basically creating a scalable Dynamic topic model I mean I Loosely call it's topic model it's not exactly a topic model um but something that is a scalable um but also reactive because things start emerging as stories break but you also don't want it to be unstable so I think we've striked a good good balance between something that's agile and it learns from the data as the data goes by um so yeah then you look at the Articles she's written on the left you know things and you kind of kind of detect those phrases big data so for as you might guess we basically map it to using our topic model to a little bit more compact uh space where we actually rank people in so the other interesting question is what are the set of topics you actually rank people in turns out that's not very trivial either but but having a taxonomy play things that we rank people is another thing that that we had to solve for so so you kind of now know that all stories she's written when you project it in the in the space of things topics we rank them that has a pretty nice Peak and she's probably writing about data science a lot um and then those are things that are a little bit easier right uh I mean easier scientifically but we had to build all these from scratch we track about nine different social media sites for every URL that anybody has published so we kind of have a Time series of how content gets shared um in all those different Topics by all those different authors uh and then the other thing that sort of interesting is um admittedly we're not there yet but like if you write for Tech crunch because you get distribution you're more likely to get a large absolute number of um social media engagements but that doesn't mean that your article is actually good necessarily it's really that rati of all of 100 people who read your article how many of them shared it um is what we're trying to approximate so it's not quite trivial um but we've we've come a long way on that too uh and at the end is like quality of article I I want to measure of how good this article is in a you know when you look at it is it something that's appetizing and you want to read it um you know I'll give you a couple examples of what I mean by that so like this uh so this is like this little professor calculator unboxing video is hilarious it might be but it's not really something I'm looking for uh which it it has entertainment value so it has value but it's takes probably about 13 seconds for an average people for people to engage with it whereas something that's more in depth and and of a higher quality so we've started building models that tries from the composition of the HTML like the images and obviously the text that goes inside it to try to figure out what a quality ranking for that article so I want things and part of that quality ranking is that opinion uh that I was telling you about right now again admittedly our opinion extraction is is rather simplistic but knowing that this article starts with building a case uh even how many images you have how big those images are um kind of are predictive of whether somebody actually ends up reading again what I'm trying to predict is whether somebody would read this thing um at this point um so opinion versus sentiment versus a statement of fact is what we're playing around with uh again I'm looking for for opiniated sentences something that when I read about it I know what I I would I would have something whether I agree or disagree with um and those are usually sentimental but they're not always uh so I have three different types of sentences here so although some might argue that pure Android is limiting I find it to be cleaner so that's good I mean it's a good sentiment and I find it to be sort of gives it away that it's an opinion of the author second one is some statement the fact third one is a little bit more subtle the second generation Moto X builds in it predecessor in all the right ways um so it's uh positive it's probably an opinion of the of the author um but there's nothing that really gives it away as a you know I believe or I find it it turns out in my opinion that type of thing it's there's nothing obvious that gives it away so just playing around with those as well um this is just you know you have to have a slide with boxes and all right um yeah so this is sort of an overall just give you a set of the feel for the scale that we're looking at we crawl about 12 million feet about 50 million Pages a day but only about two to three 100,000 of them passes our filters to really cut down on the noise uh that we see in the web um we do track about a billion social media signals on nine different sites and all of that basically feeds into our expert ranking that has that topical model um and uh and social media engagement data and basically the NLP in text analysis and quality factors through the pages um and I'll stop here thank you question guys yes what kind ofel are you using so we built our own which is so again I Loosely call it a it's it's a graph it's a topic graph that tells you what's related to what with what strength so it's not exactly a topic model in this so we do a graph clustering on top of that to have an actual topic hierarchical topic model but it's essentially a correlation Matrix that gets computed very efficiently as the new data comes in it gets updated and I saw a talk that's that's a little bit later uh about how they do that in a streaming sense with uh with spark we do that a little bit differently but it's basically an incremental correlation Matrix that gets updated once in a while we so so then you could think of it as a as a sparse graph we do a community finding on top of that to have actual topics um but we rarely use that we just use that for show off but it's really that correlation Matrix that matters and tell us how things are related to when same person appear in multiple social media do you just on page that link other social media or you try toate people they we we do that and thanks for thanks for bringing that so just like core to this is for me to know all the things you've written about we tried resolve as much as it by resolving your Twitter handle and whatnot and there comes a point where we give you get them guess that based on the topics you because we're really good at you know contextual relationships so it's how we can kind of try to say well this is most likely at 80% probability the same person how we use that it really depends on different things but but we try to have a UniFi profile of an author across SES thank you