[MINI] Leakage

Key Takeaways

[Music] data skeptic mini episodes provide highlevel descriptions of key concepts related to data science and skepticism today's topic is [Music] leakage so we bought a house we closed escrow we moved in and we moved in last Friday before jump in any thoughts on the house I tried to take a shower I came to the conclusion after the water running for 10 minutes maybe even more and the water was not getting any warmer that the water heater was broken then I went to tell Kyle here and he was like I don't think he believed me at first he was like he was like well it takes a while for the water to get warm and I was like I know I took a shower before and I know how long it takes and I went to the kitchen and ran the water too and it didn't work yeah and then even after that I don't even think Kyle really believe me but no I tried it and I believed you but then trust but verify so then we called the water heater person yeah and $65 later we learned that there's a very simple solution that all gas meters in California have a little special trigger thing that's uh goes off if it thinks there's an earthquake to stop the gas so you gave me some tasks to accomplish when we moved into the house right did I fulfill your expectations in the things I accomplished I don't know you said you fix the leak under the sink but we're still validating that one yeah how are we going to test it I'll just look randomly yeah so I you could report back maybe but that's a nice little segue for US Today's topic is leakage Linda what does uh what does leakage mean to you what's what's your sort of definition working definition uh something that cannot hold water and therefore lets the water come out cannot hold water lets the water come out or any liquid but that that's like a funnel and a funnel we wouldn't say it leaks well if the intention is for it not to leak yeah if it does if the intention is for it to be waterproof or non- penetrable by water one way we could say it is that a leak is something that we intend for things to stay in an area or to stay restricted and that doesn't happen right we we intend for the water to stay in the pipe but it gets out and that's a leak so it's supposed to be a barrier huh yeah well yeah a barrier or just a a segregation sort of thing so leakage is an important topic in machine learning actually so what does it mean the technical definition is something like if some unintended additional information that goes into your training data that your model can learn from in order to make predictions in a manner that isn't useful in practice is that a helpful definition for you now what is it in Lay person's terms in machine learning you want to train a system to make like predictions or do classifications on a given data set so you provide it some data and you say try and learn the pattern here right but if the pattern contains the answer then it's sort of trivial it's like a short circuit the system can just learn the pattern directly cuz it's already there in the data but it's like a cheat mhm let's jump over like an e commer example right okay let's say you were looking at the way people navigated a website and you kept track of like if they visited certain pages like did they visit the about us page and did they visit the blog page and did they look at the products page and you wanted to predict are these customers that cancel their subscription if they happen to visit the page that's called cancel account. HTML that would almost certainly correlate very highly with predicting that that customer turned out right mhm cuz if you to that page you seem to have the intent of cancelling your account but it's also sort of useless because no one casually goes to the cancel account page like let's just find out what it takes let's let's let's explore the website and its great design on its cancel page like if you go there the immediately next thing you do is cancel your account pretty much right yeah and like maybe you have a winback strategy but more or less if you visit that page while yes it might predict that 1 second later you're going to cancel your account that's not really useful it's not a predictive feature it doesn't help you you can't take advantage of it this is an example of leakage meaning what is leaking what the thing you're trying to predict like is someone going to cancel their account is embedded in the training data you provide and which training data would you be providing you'd be providing that they visited the cancel account page okay now on the other hand that's not like totally bad like you could say like well how many times has this person done a search on our website and that could be useful right because if someone does a lot of searches they're probably engaged and they're like you know buying stuff they're a good customer whatever and someone that does infrequent searches might be likely to say like oh I don't really use this I'm just going to cancel it so that is sort of okay and there's there's some predictive power there's good information in that but this one little place if you include the page where they go to cancel then we would label that leakage because that event is not really useful to you and it's not generally available in your training set but it's it's highly correlated with like a sequential step of the next thing that happens as you cancel so sequential is a key term yeah sequential matters a lot also like in predicting the future is a problem so generally you know you train an algorithm on historical data but you want to usually use it to learn something like make a prediction or a guess about the future since you have all the historical data if you include something that from your reference point is sort of in the future then it qualifies as leakage I guess the key way to say this is is the data available operationally meaning at the time you want to use the data would you have it available so like including information about the future like let's say you want to predict whether someone would be granted a mortgage or not and you were going to look at how many credit checks did they have in a particular calendar year okay so if you have all this historical data and you can be like well how many credit checks did a certain person get in the year 2014 and that's in your your past data you can look at it but if they were trying to get the mortgage in July and you look at the full year of credit checks it's not really representative because if you actually were trying to use this in practice if they were trying to get the morgage in July you can't look into the future because that hasn't happened yet okay why might credit checks be predictive of if you get a mortgage or not well you have to get a loan and if you don't get the loan maybe you just go and get another apartment in which case one or more landlords are going to run a credit check on you so maybe like having more credit checks could indicate someone who failed to get a mortgage and that would be a feature that a machine learning algorithm could pick up and take advantage of but that information technically is sort of in the future if you aggregate by a calendar year so that's a example of leakage when something about the future slips into your data set so I guess the key idea here is if the data is operating a Al unavailable so there's actually a famous example of leakage too that there was uh somebody was working on I think cancer rates and they wanted to predict whether or not someone would survive every patient was sort of anonymized with a hospital code but the code contained information about the facility they were at and some facilities were specialized in really really bad types of cancer that have very high mortality rates and other facilities just took lots of General patients so the facility that took the people that had the really bad cancers with very low chance of survival by definition they specialize in people that have really bad odds against them right so the fact that you went there doesn't mean that Hospital's bad it just means that they're kind of like a last ditch worst case effort kind of place and that data set contained some information that implied the facility they went to so by including that whether you know the person creating the algorithm intended to or not they gave the machine learning algorithm the opportunity to latch onto something that's not actually useful because predicting the mortality rate of someone who goes to the very specialized Clinic that's like a last resort kind of place if you already know they're going there you already know that they're in a bad situation so that's leakage so leakage is bad yeah leakage is very bad so it's wrong because a person who architected the solution didn't know that or didn't think it through we could classify leage in two ways in a lot lot of competitions where people are like competing for who gets the highest score leakage is generally used as a cheat like someone finds a way to like get a high score in like a machine learning competition but leakage happens a lot accidentally a very well-intentioned person might be trying to create a useful algorithm and if they train it and the training enables the algorithm to get its result in a way that's not actually useful or not generalizable then that's really bad form of leakage cuz then their model is not helpful so leakage if you have a model and it has leakage then it means your model is not a good model got to get rid of it got to fix it doesn't generalize okay so then I have a question if you have all these problems you know trying to build an algorithm and M machine learning mhm how do we know that we could trust you people how do you know that all our models don't have leakage in them or all your models are are flawed maybe that that could be well just keep that on the DL we don't want that fact getting out but no how can we avoid it so someone doing this type of work should be very worried about leakage and they should try and remove any field that might be related to leakage so for example like something I've done a fair amount in my life is look at the way people navigate websites and say like well this person you know goes to these pages with these frequency and they have these nav paths can I use that to be predict but in the example of like well if I want to predict that they're going to cancel their account and they go to the cancel account at HTML page that's definitely leakage and I just got to take that out of there I might even have to take out when they go to visit my account page because even though like yeah maybe they just want to change their password they might go to my account which contains cancel account also going to my account could be a good thing like maybe you go to my account regularly cuz you change your password a lot and you want to be like safety first or whatever there's no necessarily always work solution these are just some of the problems that make ml a little difficult it's not really about the algorithms it's about the knowledge representation most of the time you can also add noise that's another technique I found useful once or twice where you add a little bit of noise to make sure that whatever might be leakage is kind of like smoothed out but uh I mean these are the the challenges of a working data scientist and and for the rest of anyone who's given that model come up with a way to empirically test it weak is is bad yes indeed bad for data scientists and bad for kining well said and thank you again for joining me and until next time I want to remind everyone to keep thinking skeptically of and with data for more on this episode visit datas skeptic.com if you enjoyed the show please give us a review on iTunes or Stitcher