How Can I Sign Hawaii Banking Word

Industry sign banking hawaii business letter template secure

welcome everyone to our data science friday seminar series i'm delighted to introduce our speaker glenn nassinger i'm going to just say a few things about his abstract and his background and then turn it over to him so his seminar today is entitled process query systems with applications in security and finance and this is um uh part of this is about pattern detection problems um which in many ways the the problem is the recognition of high level behavior in large and noisy data sets and he'll be talking about bayesian statistical methods hidden markov models and machine learning techniques as tools for such problems i'm going to cut this sharp because someone's cutting a tree in my backyard and it's just so flat i mean in my neighbor's yard um dr knoxville is a data science at truest financial corporation an american bank holding company headquartered in charlotte north carolina um previously he worked as a data scientist at first hawaiian bank right here in honolulu as a member of the risk analysis division and before entering that position he spent several years as a principal scientist with bae systems in honolulu developing image processing computer vision and artificial intelligence systems his work centers on behavior detection using probabilistic methods as opposed to rule-based and he has designed applications in the district fields of sensor networks social network analysis image processing and quantitative finance he has a bachelor's in science in physics from the college of william mary and a phd in engineering sciences from dartmouth college and i i was chatting with um with dr gnostiner with one earlier about his career path so please feel free to ask questions at the end if you want to discuss you know just how did he make his way through into the career that he has now all right well thank you so much for joining us and take it away really excited to hear your talk great well um happy aloha friday everyone it's friday night over here in virginia we've just gotten through our third ice storm and i miss hawaii quite a lot so this is great to have a virtual visit um i lived in hawaii for a number of years and i i still consider it my my home so i'm very honored to be here today and uh oh i was gonna tell a funny story so i moved back about a year ago from hawaii and i started a new job at a bank here and my first week of work you know it's still going through that cultural transition and i remember talking to uh one of my co-workers and explaining you know it's friday so in hawaii you know we take you a little easy it's a little bit slower um that's called aloha friday and then she looked to me puzzled and she she just said it's not like that around here so that that was um yeah that was my uh i'm trying to spread as much aloha as i can uh back here okay so here's what i'm gonna uh talk about today um these are the four four topics and sort of four questions that hopefully you will be able to answer by the end of this talk um the first one is uh what is what is a process query system what is a hidden markov model that's a little diagram of a very simple markov model over on the right it's a discrete state model the second question is how would you use this in a security application how would you detect normal versus abnormal behavior and how would you answer a question like where did this anomaly originate so i part of my phd work was on chemical plume tracking if you had a chemical attack in the city how would you locate where it came from and then it wasn't just regular smog third question um into finance so say you're working for a hedge fund or someone involved in the financial markets or a government regulator of these markets how would you detect normal behavior from abnormal behavior and who did it and then finally i'm going to talk a little bit about a current problem i'm working on which is consumer behavior prediction uh so if you're working at a place like a bank you might be interested in a question like who will default and when or who will buy this product and when and then i'm going to try to give a little python demo there's a star there because of murphy's law demos just have a high probability of failure we'll see how it goes okay so what is a process query system this is a term created by my advisor his name is george savanko professor cybenko at dartmouth is is famous for something called the universal approximation theorem he's a mathematician from princeton and he was able to prove back in the 90s with a fairly famous paper that you can essentially approximate a continuous function with a single area hidden neural network and that's kind of a lot of the basis of what we're seeing today with deep learning and and those deeper but george started thinking what really needs to be done and and what are the in the problem we have today he was making the analogy of where we stand today with databases is sort of where we were before we had sql so he calls this the aristotelian worldview is our current approach to data so we make relationship queries such as and or if then a member of not a member of so similar to aristotelian logic where animal where you had a chart of the animal kingdom phylum uh there were strict logical relationships um so if you write a sql query or a structured data query these are the same types of queries that you make to your data say show me all uh all packets that were larger than size x that occurred between time a and b for example but what he thought was missing was a probabilistic query so imagine instead of querying um for a rule-based query like what i just mentioned imagine a query that said um show me all phone calls i received today that had a probability of being spam greater than 90 percent maybe that's not a a good one um any type of um so after going above the aristotelian world view imagine you have um george called this a new newtonian view so what did newton add to the aristotelian worldview added calculus uh so going from simple rule based uh to the mechanics of calculus you could then not just make relationships but you can make statistical arguments and queries so that's kind of the whole idea of a process to query system is to detect behavior not using simple rule based but rather probabilistic answers can be given back about behavior and it's based on something called the hidden markov model so we have here this very poorly drawn diagram by myself that shows a two-state discrete model and uh imagine you have state one and state two so if you hear the term discrete model it's really just it's just a system where you can express it as a finite number of states so for example languages are often modeled with a hidden markov model we'll get into the hidden part in just a minute but a markov model has a finite number of states and they produce a sequence so in this example we have a b a b b is a set of observations that this discrete model could generate and this is the way your phone your siri or your alexa actually takes a sequence of observed syllables and those syllables are discrete states in a language and they have certain probabilities of occurring before or after each other given as probabilities of that language so for example english and that is how um that is how voice recognition works it observes a sequence it absorbs a probability and then it generates the most likely um word that you just said okay so that's at a high level what a process query system is and what a hidden markov model is and these are the the wide range of applications that uh this has been used in so this comes from a paper from ieee computer called process query systems if you're interested um i'm sure you can pull this this up from 2007. and uh professor sevinka's students have kind of gone out into the world some in academia some some probably the majority in industry they work at places like google and one of his postdocs his name is vincent burke took this into computer security and his his company was acquired by oracle and they realized that the power for looking at computer attacks um so the problem with rule-based computer security is that the number of lines of code in an operating system are growing exponentially uh the complexity of the systems we use are growing and complexity faster than we can search them for so using rule-based um for example um looking for malware and looking for exact signatures in a brute force manner is not scalable and so you have to find ways to look at not not pattern matching but behavior detection so in the case of computer networks what does normal behavior look like people surfing the web people checking their email how does that look different than a worm that is port scanning your network so that's that's the application of computer security um because i was a physics student in a computer engineering department i got assigned the physics application um to this problem so uh we had a grant from department of homeland security at the time and we were concerned about chemical attacks in cities and so i was given the challenge of okay how do you develop a process query system to detect chemical attacks in a city so we're going to look at that in a couple slides after we go just a little bit of math so this is a more mathematical version of more concrete of a hidden market model so that the top figure here is a markov model so we have our states omega 1 omega 2 omega 3. and those are the transition probabilities they sum to one for a given node um anyone you know this is very popular if you take a computer science class you may see a weather example is a classic one you know what's the probability of today is runny rainy tomorrow's rainy of course if you live in hawaii the probability of the weather being good tomorrow is like pretty much 100 percent but i always thought the weatherman had an easy job there except for the years where there were hurricanes [Laughter] or tsunamis um so so once we take that basic markov model at the top um and we go down one and you'll you'll see the red arrows coming off so those are called emission probabilities so that's the hidden aspect so the idea is that the white circles are your states of your system but you may not be able to perfectly observe your system you may for example in the case of language if someone says t it may actually be a d so there's a certain probability of observing that sound of it being a t or a d and once we adapt this into computing and programming there are three fundamental problems that you solve with hidden markup models these are the the classic three problems the first problem is called the evaluation problem which is um what is the probability that the model produces a sequence of these states uh so so in that case you observe a sequence and you say what's the probability that the model generated the sequence so you could have a model for the weather in honolulu you could have another model for the weather in washington dc near where i live now you could give a set of uh weather and say it was 30 degrees today and it'd say well the probability of that coming from honolulu is practically zero so it's coming from washington dc so that would you could build a weather model for different cities for example that's the evaluation problem the decoding problem is given a sequence find the most probable sequence of hidden states and that's seen you have observed some weather and you want to know uh what were the hidden states that that produced that weather now this algorithm is very famous um it's called the viterbi algorithm is typically used to solve the decoding problem and viterbi the viterbi school of engineering is at the university of southern california and what viterbi was famous for he used this for decoding weak signals and this was the basis of something called cdma so if you have a cell phone produced by sprint and a couple other companies use that standard that's what it's doing so your your your hidden sequence is a noisy signal so to take your original signal you add some some noise on it now your signal is hidden can you infer what was the hidden state the hidden state was what was the original signal that was actually sent and so that's how the viterbi algorithm is used to decode noise that gets added into signals he was very successful with that and next and this is the hardest problem it's called the learning problem so given a set of observations how do we learn those probabilities and like a lot of problems in computer science this boils down to an optimization problem so we're trying to we're trying to determine what's how do we maximize the probability of this observed sequence given that's the joint probability given that we have a set of parameters so then we're trying to learn those parameters given given our set of observations so this is done with an algorithm called the baum welsh algorithm also expectation maximization and that that is uh one of the harder that's a challenging problem to solve but fortunately there are some good versions out there and there's i'm going to talk about a package you can use in python that solves this problem so just to make a little more fun i came with this example i apologize for the really um poor graphics but i had to grab some quick free graphics today so so i was trying to think of an example for working remotely since a lot of us are working remotely these days you have a you have a colleague bob that's bob over in the upper right hand corner and he lives in some far away place and you work with bob and you notice he has some mood swings so some days he's happy and other days he's kind of grumpy uh well it turns out bob is really affected by the weather you know he he has like a seasonal affective disorder or something and so you can infer given your interactions with bob that uh what the weather is but you can't directly observe the weather because because he's remote he's a remote person so you don't really know what the weather is where he's working so to translate those three different problems into this problem so the decoding problem would be given a set of moods that you observed this week and bob what was the weather in his city this week so um and then the second problem would be the evaluation problem so uh what's the probability uh that he's happy all five days in a week and you know you would you would build this up over time you would be able to build a model for that and then finally the the bound welsh algorithm how do we build a model uh to represent bob we want to train those parameters so that we can we can have a model to to represent him and once we build a model for him we could actually build a model for different people that we work with and then identify which person it is we're we're working with so so this this next chart just shows what that model would look like so so his moods we can observe so that that's the observed state here uh the hidden state is the weather in this case um so what's what's unique about hidden markup models compared to other statistical modeling methods is when you know if you look at something like regression or time series analysis that's used in say traditional econometrics we're really just looking at the statistics of the observations what's fundamentally different about hidden markov models is we're actually trying to structurally represent the underlying system and as opposed to just the statistics of the observations so i think that's kind of neat so obviously this is a very simple weather model you would probably have more than a two-state weather model but this uh produces some pretty good results which which i will show into this is a super simple example but i just wanted to give something a little fun that would kind of explain those terms a little bit better okay so these are some of the the problems that we got us we got us assigned in our in our graduate group to solve uh using uh we built an engine that was solving this hidden marker models but we wanted to show that it could be applied to different problems so the first problem was um a physical security camera so we had a camera it was an infrared and we wanted to detect and build models for a normal person walking down the street versus a thief or someone trying to break in a building and so what we would do is we would have an initial layer doing some image processing where we would detect faces and hands and things like that and those would become tracks those would become observations and then we ended up building a detector for a normal person versus a thief so the idea would be you're detecting that behavior in the video stream uh someone made the point unfortunately what we probably ended up building a model for was not a thief but a graduate student pretending to be a thief was was the model we ended up with but it did work out uh the next example was uh we had some fish in a tank and we wanted to build behavior models for the different things that fish do and fish do three or four different things when they're living in a tank they they they eat they sleep they fight and that's pretty much it so we we basically built models for each of those three or four different activities and then we would have the the camera observing and this was only two-dimensional so you know it was it was a little bit crude but we had a two-dimensional movement model basically and then it at each point of time those models that had been trained would return a likelihood saying the likelihood of fighting is you know 80 the likelihood of sleeping is 20 so then we we would in real time get a get a a probabilistic number for what are the fish doing right now unfortunately this led to all of the graduate students in the lab having to go through animal training because they considered a webcam on a fish tank is animal experimentation so uh that involves a lot of paperwork uh i got that job so we had to certify that we were we had a guillotine for humane disposal at end of life and stuff like that okay so uh this is the problem i got assigned which is the the chemical problem since i was the physics guy and i wanted to do a little bit of a thought experience so imagine you're in a city and you have binary sensors that are in your city kind of like a smoke alarm and they're set to go off when they sense some kind of harmful agent and something gets released it could either be a fire or it could be an actual attack of some kind and the sensors start going off spatially in some kind of a pattern and then you have to answer these questions where did it happen how many sources were there and are these different observations correlated are they independent so this is what it would look like in a really cheesy matlab simulation so you start the blue dots are sensors and they're in xy space and then i'm going to step through time so time zero there's no detections everything is clear and then at time one oh we get some there and there and then at time two we got some there time three there time for there so then the question is you know what happened is it coming from one place is it coming from two places so the the traditional sort of fluid mechanics um atmospheric science way of solving this is is would be you know doing an analytical solution and this is this is the diffusion equation in two dimensions which you can solve using differential equations and you could work it backwards and um this does not really work if you have five dollar sensors that are you know uh operating the thousand thousands of them across the city so we were looking for a a shortcut way to do this without having to solve these you know really complex sets of differential equations and so we applied the the process query systems and there's kind of different approaches to how you would solve this um imagine on one axis you have your complexity of your model and then on the other axis you have the mobility of your sensors so some people are proposed to solve this problem with like swarm robots so these are like little robots that you know they either swim around the ocean or they drive around the city and they're constantly looking or smelling and then in the upper right hand corner you would have just a few very sophisticated robots and so the idea is you have a lot of really simple things or do you have just a few really complex ones and then the bottom would be rather than stationary you have things that are stationary rather than moving around because they take less energy so the approach that i took was this approach 3 which is static sensors that are really cheap it turns out they use a lot less power so in robotics 90 of the energy you're going to spend is for motion moving or flying and so if you can have something be stationary and sit there the power consumption can be much much less so it can that's why your smoke alarm on a 9-volt battery can run for like a couple of years um so basically what we did is we if we have any uh bayesian enthusiasts out there we we applied bayes rule uh which is a conditional probability so you're essentially taking that diffusion equation that i showed you before and you're flipping it uh you're inverting it so rather than predicting how particles are going to disperse forward in time you say given an observation what's the likelihood that it came from this area sort of like the inverse of remember when the hurricanes are coming towards oahu and they're showing the inc forward in time the the area of prediction gets larger and larger the same thing happens when you look backwards in time so if you observe something and you look back a short period of time there's a small area it could have come from if you if you go back further in time it creates a larger tone of uncertainty where it could have originated so we take these likelihood functions they go backwards in time and we overlay them so we take our observations like at the top we overlay them they create a likelihood of where the stuff could have come from and that then we can create queries um to estimate where things came from i'm not going to go i don't have time unfortunately this this would be like two or three entire other talks but i'm just going to mention a really cool concept that you might want to look at that it's based on it's called multiple hypothesis tracking and i i just think this is a really cool idea um it originated back in the 60s and it was used for airline or radar tracking so imagine you have a set of observations and you have to explain what generated them so using something called common filtering if we detect say a 747 we can make certain inferences about its possible movements whereas if it's a a hang glider it's going to have a different set of potential observations between time points so these a b and c possibilities show that given a set of observations we could explain in three different ways a is that there's two airplanes that just traveling in a straight line possibility b is that there's two airplanes traveling in a straight line but that third measurement is kind of an error it's just a little bit of an outlier but it's still just two tracks and then possibility c is you have two airplanes but they actually swap tracks and you're now measuring the um the two different airplanes have swap position so that multiple hypothesis tracking is sort of at the core of what we were just talking about because you're maintaining in this case you have three different hypotheses that can explain the same set of observations when as new observations arise you can now rank them in likelihood and you can you also need to do something called pruning because uh it's a combinatorial explosion over time you end up with more hypotheses then you can continue to maintain but this is part of the whole pqs concept we maintain multiple hypotheses to explain what we've seen and we update those likelihoods as new information arrives okay and this is uh just more more kind of simulations um i ended up doing my phd work in in computer simulation because there was a lot of paperwork involved with trying to release chemicals uh with humans and things like that so i learned to code more that was that was the upside uh once again this this shows sensor observations that they receive and then inverting that into a likelihood of what produced these observations same problem that we saw with um the speech recognition just applied to a different area okay let's get out of physics and let's get into something else so i did not plan on working in finance it just sort of happened um i was i was working out in hawaii i was working for uh for an aerospace company called ba systems and around around 2008 we had this thing called the financial crisis where a lot of a lot of companies almost went bankrupt and a lot of financial systems were teetering on insolvency and so after the financial crisis the federal reserve and most regulators around the world got really serious really quickly about running simulations under different scenarios to make sure that your institution would survive if there was another crash this was called stress testing so it turned out that there was a huge shortage of people who knew the finance and knew the math so they the group that i was at at first hawaiian bank we had oceanographers we had computational biologists we had mathematicians we had computer scientists that were hired and they just taught us the finance part uh because we knew we knew how to program we knew math and um it was a it was really fun it was it it was uh it was a very serious problem and there was a lot of resources applied to it and so they built up a group very quickly and the whole stress testing thing was a really big deal for three or four years and uh it still is still is um but that that it was a great opportunity um to to learn about financial markets and to to combine the the engineering and computer science with finance and so i didn't plan it it just sort of happened and it was an opportunity and i and i took it so it was a good entry um but there was another graduate student in my same group that ended up going into finance his name is tardowski and he's currently in california working for a pretty famous fund called dimensional fund advisors and his his phd thesis was really interesting so i don't know if any of you are into to stock markets or how stock markets work there's something called a market microstructure so if you go home and and you want to buy some gamestop stock and you you put in an order to buy 100 shares of gamestop what happens is it gets routed to what's called a market maker so that uh there's a handful of these firms and they're the ones that actually place them this is an actual picture of the new york stock exchange here so it's basically a server room in new jersey is what it is and if you have the resources your trading your market maker will be located very close to this machine uh maybe across the street if you're very fortunate because you are limited by the rules of physics and the speed of light and you want your orders to be a few nanoseconds faster than your competitor um so what what twardowski was able to show was that he could d anonymize trades so he built process processes to represent different types of traders now it turns out that your trades if you if you place a trade for 100 shares of gamestop your trade is anonymized to everyone except the government regulator so because he was working with the government regulators he had he had access to to this um this information and he was able to to to get ground truth so so if you were to simply look at all these orders that are placed so these green dots and red dots show orders that are entered and removed most orders never actually go through they're placed and removed placed and removed placed and removed he was able to look at the timing of the orders that were placed using that information in the structure contained in that he could say oh that's goldman sachs putting an order for 10 000 shares uh i'm totally making that up um that's an arbitrary name it could be x i'll say xyz okay um so you know this has applications in both um finance as well as computer security so given time you know very very my new time signature information he was able to de-anonymize cluster behaviors and say that the actions that happened at this point in time were the same actions that happened at this other point in time by the same actor and so he he built these models and he this was the observation sequence he was able to assign likelihoods uh to his models based on the set of observations so it was pretty cool so the one of the problems i ended up working on was something there's a professor kleinberg at cornell who wrote an algorithm called the burst algorithm and he was interested in the problem of he receives all these emails during different points of the year uh he received letters about letter of intent from students um he received proposal deadlines you know he had notification emails so he was just swamped with emails and he wanted to try to figure out if he could develop a chronological temporal information that could then be used to classify the type of email it was so for example if it comes in at a certain time of year there are these bursts now the interesting thing about bursts is they occur in email they also occur in things like in astronomy like pulsars they also occur in stock markets and so um i don't if you're interested in his work it's super interesting i highly recommend you checking it out um but i don't have time to go into it right now i'll just mention that his his ideas for classifying email uh we were able to apply uh in financial markets uh so the idea of temporal burst structure and this this is one this is one problem that i'm gonna show as a burst so in 2012 there was and i emphasized was because they went bankrupt after this date there was a a company called knight capital remember a few minutes ago i mentioned there was about five big players well they were one of those they had some code that accidentally put millions of orders in at the market open on august 1. [Music] so you may have any of you that worked in finance there's something called a market order which means i want to buy the stock and i'll pay whatever it takes i don't care what the price is just give it to me and then there's something called a limit order which is i will sell it but i will only sell it at this price so there was an error in their code it's actually documented the sec did an investigation you can go back and you can read it it's pretty fascinating about how mistakes in your code can lead it's like listed in the top 10 massive failures due to computer code they ended up losing billions of dollars that morning sorry 460 million dollars that morning the brokers who were using the software received this cryptic message on their screen that said power peg disabled and they were like power pegged is it what what does that mean it was it was obscure message that didn't really explain what was going on and this event generated you know stocks like coca-cola were going from ten dollars to forty dollars within a few minutes and and those were the stocks that were on their buy list there was other stocks on their sell list so there were stocks that were going from you know eighty dollars to five dollars in the same period of time and so the sec came in a lot of the trades were busted the the in most of the individual investors got the trades cancelled and got their money back but the company went bankrupt but what it generated was they burst and so it was a great data set to work on um similar to a pulsar similar to a kinematic explosion it it was something that nothing's happening and all of a sudden something big happens and and so how do you how do you how do you detect that how do you how do you differentiate normal behavior from something that you've never seen before that happens and so that this is the type of behavior that you would like to be able to detect and deconstruct and then pinpoint who did what when okay moving along in finance um another really interesting problem uh it's called adverse selection i love this problem um it was actually resulted in a nobel prize in 1970 called the um george akrilov so he was looking at information asymmetry so imagine you're in a market where uh one side has all the information and you don't uh so for example use car market the guy selling the car knows all these problems exist but you don't and therefore that's asymmetric and this this creates um imbalances and it ends up hurting veryone because no one trusts anyone um and the reason that the bank cares about adverse selection is the following so it sounds really boring at first your job is to price a loan so people are coming to you and they want to borrow money but you don't know how much to charge them so you're like well okay i want to be a nice guy i want to charge everyone the same i want to make it equal so i'm going to charge everyone the same sounds great but here's the problem if you charge everyone the same what happens is the really risky people who would have cost more they're going to flock to you because they're going to say oh wow this is a great deal and then the people who are really you know very responsible and low risk they're going to be like this is expensive it's not worth it to me i'm out of here and so you're going to end up with an entire portfolio of really risky people and you're going to go bankrupt and so the the problem of adverse selection is appropriately pricing risk and that's that's essentially what banks do is they price risk and so if your competitor is better at pricing risk than you are then they're going to win because they're going to get better prices to the lower risk and they're going to inappropriately charge people that they shouldn't and so that that's the idea is to appropriately price risk and so in behavior modeling that's that's a big part of what we do is given you know given behavior how do we estimate the probability of default given a certain history now there are there's there's lots of regulations about what information you can and cannot use to make these decisions but within within the parameters of the information you have available it puts you on the same playing field as your competitors so if you're able to identify factors that are still within the legal framework of what you are able to use then then you it's win-win because you're you're giving someone a good deal that deserves it and you're lowering the risk and you're lowering the price for them and so that's uh and building trust and so that's uh part of what what we do at a bank uh is adverse selection and default modeling and the way this falls into hidden markup models is following um if you were doing the same thing applies to corporations as would apply to individuals you want to figure out what companies have a high chance of making it what chan what companies might not make it if there's a downturn so the way companies get rated is his following triple a double a and a it's kind of like your fica credit score but it's for companies it's the same thing and so by building a markov model you can say okay given that this company and then you can get really fancy you can build something called a conditional markov model so given that the this person is currently a triple a and given that unemployment goes to 12 and given that gdp drops by 10 what's the probability they're going to go to single a so you start building up these conditional conditional probabilities and transitioning [Music] this is also used for modeling the the probability that someone's going to you know one day be a really great customer and then you know one quarter from now they're gonna they're gonna max out their credit card and disappear okay so now we get into the uh the customer behavior application and in the demo that will hopefully work um so in in customer behavior work uh and this applies to and i'm giving you all some some really golden nuggets here because with these tools you too could build a process query like system it's really in in two of the three parts actually are already available in free open source packages so the first part of your problem is you want to isolate your processes so what i mean by that is remember when we were looking at the the fish and we had fish that were fighting and fish they were sleeping and fish they were eating so so you that's called manual labeling so you would actually have to wait until the fish are sleeping like catch some video of them sleeping and then go train your model okay so that's that's manual labeling and that that works in a lot of cases if there aren't too many different uh processes but ideally and i've become more and more convinced of this in my um industry work um you really want to go for i'm a big fan of unsupervised learning if you can pull it off so as an example currently i'm working at a bank we have hundreds of thousands of customers we want to find the different types of behavior that they exhibit there's no way that i can go look through 300 000 accounts and try to identify examples of all the types of behavior i mean i suppose if if we if i had like a thousand interns maybe we could pull it off but um we really want an unsupervised way to do this and and so i've become a huge fan of using unsupervised learning uh your those of you in computer science i'm sure you're familiar with things like k-means that's the real basic type of unsupervised clustering [Music] and by using an unsupervised algorithm you can learn the structure of the data and i'm also a big fan of that because i'm letting the data speak for itself so rather than me imposing my view on the data and saying i think there's probably four types of customers and i think they're the following i'm going to let the data run the unsupervised algorithm and it's going to tell me how many types of customers there are and so that um that's been very powerful so for time series there's this really cool package that i'm a big fan of called stumpy or stumpy and it's um this researcher his name is law law he works at td ameritrade but his his implementation of the package is actually based on uh the research from a professor at uc riverside by the name of keo or co kao gh so he published a number of papers and the stumpy package the other cool thing about stumpy is that it's been optimized to to run on gpus distributed processors so you can scale it up and what what stumpy does and i'll show you in just a minute is uh it takes time series and it finds what are called motifs so it finds sections within the time series that are like other sections in the time series and you get a distance metric so then you can do something called clustering and you can find similar groups of people or similar types of behavior in the time series i'll show you a more concrete example in a minute but basically the stumpy package can be used to take a time series and unsupervised find the different types of behavior once you've done that there's another really awesome package called hmm learn which is hidden markov model learn and build into this package are those three different algorithms that i mentioned the learning problem the viterbi algorithm so you can do all of that using hm learn so once you have your different sections of your data you can train models for each of those types and then run the detection problem to detect that behavior in your data okay so now we're gonna we're gonna get real brave and see if i can do a little little code here okay so this i thought this was a pretty neat data set this is passenger taxi data with half hour increments from new york city and it's for about two months of time so this shows the number of passengers taking a taxi in new york city so it's a time series uh so first we're gonna load up uh pandas stumpy and a couple plotting libraries and we're going to load the data set okay so now we just ran we ran the stumpy algorithm and let me explain real quick what it's doing this matrix profile is for each uh so if you take a sliding window and you move it across your original time series for each window this metric tells you the distance and it's called the standardized euclidean distance to the most similar other time point so it's it's the it's the distance so what that means if you have a spike that means at that point in time the closest thing to that thing is not very close so in other words there's something unique about this point because nothing else is close to it so it has a high distance metric from any other point and when we go back and we check the calendar and we label it well you know if you were to just look at this with the unaided eye it's a little hard it's a little hard to see you can sort of see that one spike that happens but the the other is not so much uh but when we run the anomaly detector on it it shows oh there's three anomalies and what are those anomalies oh it turns out the first anomaly is columbus day and if we go back columbus day is kind of a big deal in new york city maybe not in some other places but the traffic was down it was a holiday so there were not as many people taking a taxi at their normal time um daylight savings time there was a huge spike i guess people were running late to work i don't know um they needed maybe they normally took the cab and or they normally took the bus and they're running late so they had to call a taxi um and then thanksgiving um same thing there's a there's a spike in the in the minimum distance during thanksgiving time and you can sort of see it once you go back to the original time series you can you can kind of see it but it would be hard with to to me it would be hard to look at the original data and see that so by so by looking for for local minima and maxima you can you can do things like finding motifs finding outliers in the time series and then take these time series feed them into your training and look for other periods of time that have that same behavior this is the window size just to show you that you know a lot of these machine learning algorithms are very parameter sensitive um and to me that's something to look for that if i don't put in exactly the right parameters it's still kind of going to work so we can see here we're using a range of window sizes and it still produces the anomaly detection maybe not quite as sharply but it still mostly works so that's something to be aware of is how sensitive are your algorithms to these parameter inputs occasionally they're algorithms that you have to have everything over fit and exactly set or else it doesn't work at all so stumpy is if you're interested in time series check it out okay and then i don't have time today to go into hmm learn but basically what that will do is once you've identified your different either outliers or periods of interest you can feed that into your hmm you can learn those and then in the future when you see that again it will detect them okay we're short of time so i have one final slide where i had a few career tips that's what i called it there we go so i ended up in industry it wasn't exactly planned these are just a few little ideas i wrote down from from the last few years so the first thing i wanted to mention is it's important to ask yourself is it computable so there's a lot of really cool algorithms out there for machine learning and ai that may not be practical and especially if you're working in finance or other areas the data sets can get really big or really complex and you often have to be able to quickly do some order of in type calculations in your head can i for for the computer resources i have and for the algorithm i want to use can we do it and and you may have to take your approach and dial it back a little bit and simplify a little bit um so just as an example i said your boss may not want to buy you a gpu cluster if my boss watches this online i would love a gpu cluster um or they may not be willing to buy you a thousand hours on the amazon club so that's just a problem that i find in our research group we run into almost a daily basis like we want to solve this problem we want to use this approach can we actually do it so over time you're going to develop some more intuition about which approaches are appropriate for which data sets and that's that's a really good intuition to have so if someone comes to you and says can you solve this problem and this is the size of the data and then you say what's the size of the computer yes yes we can compute something in a few hours or no no we can't so that that's a good intuition to develop the other thing i was going to mention is if you haven't an ability or an interest in leadership there's a huge need for people with technical understanding who are also good with people who are good leaders there's a huge lack of people especially the higher up the louder you go part of it's because machine learning and data science is a new field so they're just not in the senior positions yet and i think there's a huge need um they often get taken advantage of by consultants because they just don't know and and so um i think there's a huge opportunity down the line in the next few years to take technical skills and also combined with leadership and business skills huge opportunity so i was just going to mention like if i was looking for a job what kind of job would i look for what kind of to me the team is very important so if possible i know you don't always have tons of choices but if you if you do have choices um i think it's very important to look for a collegial team a team that works together where people truly are looking out for the best interest of the team and i can brag on my team at first hawaiian bank because that's the way we were it was such a wonderful place to be because it was not a top-down um approach it was very collegial we worked together sort of an academic type approach where we were we were all colleagues we came to consensus decisions as a group uh which was really great so that's something to look out for um beware of being a loner uh some some you can sometimes end up as the only data science person by yourself and those that are in charge of managing you may not be able to correctly balance expectations so they may give unreasonable expectations or they may not understand why is it taking so long or um so it's good to to have at least one or two other colleagues to work with ideally um we discussed this a little bit at the beginning of the call i was just saying there are these cycles and kind of hype about there's something in business called the s-curve where things come into popularity and then they go out so you know in the 90s neural networks was was going to solve everything and then it gets over-promised and then people say oh well no one everybody knows they don't work and so i would just encourage you don't get too caught up on you know um getting the exact right title on your your degree or the exact right certificates um just kind of focus on the fundamentals you know programming math computer architecture networking and um you'll be fine so um i made the analogy uh so during the gold rush of california the famous story was the people who really made all the money were not the people digging the gold they were the people selling the picks and shovels so so so there's a lot of people out there selling picks and shovels to to data scientists um you know different this is who just you know focus on the fundamentals don't i wouldn't worry too much about chasing you know the certain titles um just just learn that learn the stuff and you'll be fine and then finally this is a little bit of a more fluffy advice but i have noticed that attitude is very important and the people i've met that have been very successful um they're very not just positive but they they consciously choose to to focus on the opportunities of a given situation you're always going to run into people who are going to tell you why this won't work um what's wrong with the world why this uh you know just try to keep refocusing your energy on the possibility and what's possible and you do need to hear those those warning voices sometimes like hey what you're doing is illegal you really shouldn't do that okay i'll listen to that but uh just you know try to focus on the possibilities and that is all i have and i'll be happy to entertain questions thank you glenn that was super interesting um i think so folks can either put their questions in the put their questions in the chat or just ask your question i think we have a good group here today any questions for for our speaker so before people start leaving uh our or asking questions i just wanted to give a shout out to lizzy fink who is one of our former fellows who actually su gested that uh that doctor not that we invite dr noxinger to come give a talk so that's how how that connection happened happened and so thank you lizzy absolutely okay i'm sure we have some questions here [Music] well glenn actually i have one um so i was really interested in in applying some of it the first part of your talk we're using fish tracking methods to attract people and others do you see is that you know is that kind of well accepted because you're taking something that's really kind of basic science taking those ideas and that and that valuable research but twisting it a little bit to um to apply to a different problem because sometimes it's like oh wow what is fish tracking going to tell us about anything that's going to be really important to finance to whatever but of course that's that's that's their favorite thing in the world to think about so do you have other examples or comments around so in other words um examples of how fundamental problems were adapted to applied problems yeah exactly oh too numerous to even mention um i mean i i think i mentioned you know computer security uh enviro virus detection so this the same concept uh was applied to finding computer worms um and was and was commercialized into a huge company um and uh social network uh so for example another another colleague of mine was working on applying this to i didn't get a chance to talk about the social networks so one of my friends actually he ended up working on the enron data set so i know some of you are probably too young to remember there was a company called enron back in the day who was very bad and um went bankrupt it went it was on the cover of like time magazine one year is like best investment ever and it turned out to be the whole company was a fraud and they were doing fraudulent accounting and once they collapsed and went under many americans lost a lot of their money because they had their stock invested in it and it's the best example as to if you ever worked for a company you should not put your retirement account in stock with that company because if the company goes bankrupt you not only lose your job you also lose your retirement anyway that's a different topic but my friend was because this company uh you know many of them went to jail and so forth uh they created a really interesting data set so the data set was an email um so my friend ended up working at the fbi i can't really say his name or anything like that but he was give he he did work on the on the enron data set and using these methods they were able to take uh he was able to show uh given an anonymous email stream uh over time of people sending an email to each other he could de-anonymize which people were lead ring leaders versus which people were followers versus which people like what role they played so that's another kind of behavior detection example of taking email streams and then trying to infer types of behavior i still keep hoping that google that gmail will come up with a better way to like automatically answer our emails or at least cluster them because that is i want to i want to figure out a way to stop getting um all of these uh spam phone calls on my cell phone but i think i think everyone resonate with this one of the challenging things about email is that there's no order to the messages it's here's a task here's a new task here's someone asking you a question here's something that you're late on and you've got to turn something in really quick so it's very hard to kind of have anything other than a really disjointed kind of work session when you're kind of going through your email definitely that's a problem i'm not sure we have a solution to it there now we can say lizzy see lizzy lizzy do you have a question maybe i'm so sorry my audio cut out about five minutes ago um i do not have a question but i'm really grateful to dr knopsinger for for speaking to us really enjoyed it very good well thank you for inviting me i i really enjoyed it and um if you all have any questions or you know career advice questions feel free to email me i have a website if you just google my name you'll find it and you can contact me there and i would be happy to give feedback about you know transitioning from academia to industry or data science careers whatever that would be awesome you know laura maybe that's something fun we could do with the fellows bring all the fellows back together and kind of you know for a little alumni gathering and talk with some some folks like glenn and others about career prospects i think that might be really fun yeah now in our we're now actually in our third cohort of data science fellows so we've got about you know a good group now and i think the most interesting well the data science fellows is really specifically designed to bring together um students of very different backgrounds so astronomers oceanographers some computer scientists and others you know as a cohort to kind of share the experience and see what they can learn from each other too so we we also learn a lot from those different perspectives with you know kind of data science and analytics being the common theme through fellowship so i i think i mentioned our data science team at first aquarium bank if you were to pick the single largest background it was actually oceanographers you know okay like just keep hearing things like this the other thing we often hear is that there are no unemployed astronomers two astronomers oceanographers and others so i could just tell everyone there's life after graduate school even if you didn't absolutely and honestly not everybody wants to be a professor we're we're not all smart enough that wasn't what i meant okay it was uh it was a little um there was a little bit of challenge transitioning from academia especially my first job um but then you get used to it and you know like anything else there's pros and cons mm-hmm yeah um but overall it's there's tons of opportunities and now is a very exciting time to be in data science it it changes every month like it's so fast so i think it's really exciting to be in this area right now awesome well any other questions for glenn before we take off for for our weekend okay all right well with that thank you so much this is a really fun so i'm just seeing all kinds of comments in the chat people really appreciated your talk so so um feel free to join us in time if you want to come and listen to any of our seminars and if we do have meetings with our fellows we'd love to have you back thank you so much i would like to do that thank you have a great weekend everyone okay all right aloha