Have you ever applied for a credit card? A few weeks ago I was up late at night and apparently my idea of a good time is to try to churn some credit card rewards. So I was on a bank's website entering information about myself, and I got to a point where I had to hit a submit button to submit my application and I clicked submit and within a split second the page loaded and they told me whether I got the credit card, and my mind was just blown by this because I thought, "Surely they don't have a squadron of people sitting around reviewing applications. It's 2:00 a.m.. And even if they did those people couldn't possibly make a decision that quickly. And I was just very confused by this. But the secret here to how they're able to do this lies in machine learning. In other words people can take past examples from history and use it to create a formula that allows them to predict future outcomes. I am Samuel Taylor as introduced and today we are going to be going through a crash course in machine learning. There are a lot of things that this talk is and aspires to be. And chief among them is to introduce y'all to machine learning concepts. If you don't know anything about any of this stuff I hope that you are able to come in here and feel like you took something away.
But if you do have a little bit more experience I hope that this is going to be helpful for you still because you'll be able to see some different approaches that you might not have considered before and hopefully that will be beneficial to you. I definitely want to orient this toward application because I find that sometimes when we talk about machine learning especially at an introductory level it becomes a really weird and hypothetical scenario where you're saying, "Oh well if you have this situation then you might want to do this" and it just gets weird to talk about and it's a lot more interesting to talk about an application level. That said it is still very respectful of the theory in this field because there's a lot of really talented people who've put a lot of really good effort into understanding the way that these models work. And if we are able to gain some of that understanding, we'll do better at the application. However, you're not going to be a data scientist after this. We have about 45 minutes. This is not going to be an entirely comprehensive event here. There is a lot of stuff that I won't be able to cover and a lot of things that we're going to have to gloss over because again there's only 45 minutes here. I think that tutorials and detailed code examples are really interesting and there's going to be a few code snippets but those are, I think, less applicable in a forum like this.
I'm happy to talk to any of you afterward about specifics and give you links to GitHub repos and stuff if you're interested. But for the purposes of this talk we will try to stay a little higher level and talk about what the models used were and some of the techniques and try to focus less on the specifics. This is how we're going to be doing all of this stuff. We'll start off with a few minutes just on what even is machine learning and then we have sort of a warm up use case that is this credit card application example I'm talking about. And then we'll dive into 3 use cases. Basically three problems that I ran into and the ways I ended up solving them including what I learned from those. Finally we'll just wrap up everything and hopefully tie a nice little intellectually tasty bow on this package. So for those of you here in this room if you've heard the phrase machine learning before raise your hand. OK, some of y'all liars because I have said it at least three times already, but now I know which ones of you are aren't telling the truth. If you feel like you've done something with machine learning that you found significant or just interesting, I'd love to see a hand. Awesome. OK. For all of you, I hope you get something out of this for the rest of you, we're so glad you're here, and I think that by the end of this you'll be able to try to do something really cool.
All right. There is a lot of stuff that is machine learning really and a lot of times people break it up into supervised and unsupervised problems. Again this isn't comprehensive. Supervised learning is what we're going to talk about for most of today and what we'll talk about first. In supervised learning we have a set of data and a set of output data and the input data maps to the output data. So we could have, for instance, credit card application data that maps to whether or not someone should get the credit card. So let's talk about classification first. In classification we have data that looks something like this. This is again for credit card applications, and this is entirely hypothetical data that was generated by me drawing some stuff on a graph and then reading the graph. So it's completely fake but it gives you an idea of what's happening here. In supervised machine learning we have our features and our output, and here the output is whether or not someone is given credit. That's the thing we're trying to predict, and the input (or the features) are the age and the net worth.
So we have some input and we're trying to produce some output. You could take all this data and put it on a graph and we say we have seen these six credit card applications in the past from people of various ages of various net worths and we want to know whether to give them a credit card, right? And for all of the hype and excitement around machine learning this is pretty much just drawing lines. That's all we're going to end up actually doing here. So we're going to do some really fancy line drawing today. And we can draw a line through here that separates these data points and then when a new data point comes in, we can ask the model, "Hey, should we approve this credit card?" and it can look and say, "Well, it's on the 'approve' side of the line so we'll approve this person."
Regression is very similar to classification in that it has still input data and output data. But instead of predicting what kind of thing something is (trying to predict a discrete value with a certain number of outputs), it's trying to predict a continuous value. So we're trying to answer the question "how much" of something. To continue in the bank, let's say we are trying to give people loans, and someone comes in and they say "My net worth is this, I want a loan. How big of a loan are you willing to give me?" And you could have a bunch of data that looks something like this. And we could see well we gave this person with a million dollars a very large loan and we gave a person with less money less money. And then again all this is is just fancy line drawing. So we draw this line and that's what we call "fitting" the model and then we'll want to predict a new customer that walks into our bank and they'll say, "I have $500,000, how much of a loan are you willing to give me.?" And you just jog up to the line and then jog over to the Y-axis and say we're going to give them sixty thousand dollars. So pretty simple stuff with supervised learning.
Unsupervised learning: also interesting. One of the biggest areas in it is called clustering. It's probably one of the key algorithms to know in this space. With clustering, we don't have any any outputs. We just have inputs pretty much. We see we have nine data points and here each data point is a different shape and we might have a table that has the number of sides and the color that this shape is but we don't really know what what we're looking for here. With clustering we're just trying to uncover some underlying structure in the data. So we might hand this to a computer and say hey split this into three groups for me. And it seems like you can split this into three groups a number of ways but the computer might just say here's three groups of these things. And it will provide you with those 3 clusters.
There's a lot of other stuff and this goes back to the non comprehensive aspect of this talk. This looks like a slide. It's actually secretly an insurance policy against angry people on Twitter saying that I didn't cover their favorite subject.
So all in all the way that I sort of think about machine learning is that we want to find some function in the world. There is a mathematical function that describes whether someone should be approved for a loan whether they are going to default on a loan that exists out in the world. The problem is that we can't possibly know that because there are so many factors going into it. I mean, if someone comes into an unexpected medical expense they might default on a loan and there's all sorts of different factors that affect this. And so it's impossible for us to know all of these things. The interesting thing and the thing that gives us hope is that we're able to measure some points from this. We can look at the past and see: well, we know that these points aren't a perfect representation of the actual function but we can see kind of some area around it and then build something from there.
So the goal in machine learning especially supervised learning is that we are trying to find some algorithm which will give us a G of X to approximate F of X. So we're not going to be able to find the true function, but we want to find something close to it. So all that said we're going to be walking through three use cases today and I wanted to just have a warm up one so that we can all get used to this. We're going to be walking through five questions for each of these cases: What's the problem? What does the data look like? What kind of machine learning problem is this? And then we'll dive into some details on the solution, and we'll talk about lessons learned. So in this case for our credit card application data the problem is we're trying to decide as a bank whether we should give a consumer a credit card and the data we've already seen today. It kind of looks like this. We have some input features we have output. So here is the time for all of your beautiful faces to shout at me if you're angry. Get all of your aggression out and tell me, "What kind of machine learning problem is this?" Supervised, yes. Yes. Any more specific than supervised? You're going to be like a little. Yeah. Classification yes. Y'all are brilliant, this is amazing. Yes, so this is a classification problem, and we could solve this by drawing a line because that is what we're doing today.
There's a library in Python called scikit-learn that implements a lot of these algorithms for you so you don't have to waste the time doing that yourself. And basically what you can do is import a model and say, "fit this model to the data that I have," and that will do the step of drawing a line and then you say, "predict" on a new data point, and you give it the input data, and it will give you the output that it thinks is true. That's sort of what you do. But then we kind of ask ourselves what what did we do here? Did we actually accomplish anything? Is this any good? So we want to know how accurate this thing is. In this case, obviously this is very contrived data and it's deliberately drawn to be easy. And obviously it's going to get 100 percent accuracy on this. But what does that mean, and how can we think about this more generally?
To give you an example I have taken a true function that I made up, and I put it on a graph. You can't see it yet. But then I drew some observations from that function to sort of model what happens in the real world and that's all these little blue points that you see. And then I fit two different models to them. You can see they do different things. Basically one of these is better than the other. I think you might be able to tell. Does anyone think Model B is better than model a approximating the true function underlying this? OK. OK is anyone going to be brave and say that model A is doing a better job here?
OK. OK. Some brave souls out there. So in this case the true function is actually definitely modeled better by Model B. In usual situations we don't actually know what this green line that I've drawn here which represents a true function is we just have observations. And so for a human being we can look at this and say, "OK yeah clearly Model B is better" but there's caveats of being able to tell a computer how to do this. So what we end up doing is holding out some data for testing and it looks like this is a little bit small on the screen up there. So I will describe that about 20 percent of these points are red and the rest of them are blue. And what we've done here is we had some data that came in to us and we said, "Some of this we'll call training data and some of this we'll call testing data." And what we can do is basically train our model on the training data and then predict the values for the testing data and compare the two. Then after that, ideally what we could do is calculate the actual cost of making an error. So if I was designing a system that could tell from radar data whether a nuclear warhead was headed toward the United States and I said there was one and there actually wasn't one that would be a huge mistake. Probably a lot of people would die. Maybe everybody. However if someone is able to fingerprint into my phone that is a much lower cost than literally the destruction of the human race.
It's basically going to be some dank memes are going to get out there and that's not really going to change the world that much. So ideally you'd be able to determine what kind of error you're making and how expensive it is and optimize your model for that. In the real world it's not always as clear. And so there are some error metrics that you can use to try to help you understand how your model is performing. One common metric for regression problems is mean squared error. I've drawn a little example up here with some true data and then our predictions. And basically what you do here is you subtract the true value and the predicted value and then you square that number and you sum all of that up and then divide it by the number of points are predicting for and we can say, "We were off by 80 you know the metric here is eighteen point three five" and you try to minimize that error when you're comparing models to each other.
In other cases like classification you can't use that because it doesn't make as much sense but one common metric that makes a lot of intuitive sense is classification error and basically you're just trying to see, "Did I do the right thing.?" So if on four of these points I correctly classified that they were a certain class then I would say my error was twenty five percent on those points. So in this use case we've learned a few things. The first thing I would say is that this stuff is just pretty neat. There's cool things we can do. It's less intimidating than we thought it was; it's just drawing lines. And then also that it's important to withhold testing data so that way we can evaluate how our models are doing. So in our next use case we are going to talk about teaching a computer a sign language.
So the problem here is that I don't know sign language, and I want to communicate with deaf people because they have valuable things to communicate. And I was trying to think about this problem and a friend and I were going to a hackathon when we were in university and we were thinking of what we could do with this little toy that we had. This is a thing called the Leap Motion. It's a little sensor that has an IR LED in it, and you can see that in that picture here because cameras are weird. I guess I don't know all the the weird camera stuff that makes infrared visible to this camera but not to the human eye. When you're actually looking at it in person, you can't see that. But basically what it does is it flashes infrared light up on the human hand and can tell basically where hands are in space. So for this example someone has a little sensor sitting on their desk or potentially mounted to like a VR headset and they're holding their hands kind of like this up to it and you can see what the computer is seeing here. Each of these little dots that is connected by the little pipes I guess is a place on the human hand. By and large just joints in the hand and then also fingertips. And we were looking at this and thinking OK what if we could do something with this data to try to help with this problem with sign language.
And we only had 24 hours, so we thought, "What if we limit the scope here, and we just try to do something for American sign language?" (because there are a lot of dialects of sign language). And also we just tried to do the alphabet. Let's start small and see if we can get anything working. So we gathered some data. We took our little device and plugged it into the computer and then we made an "A" above it and we said "This is an a, this is an a, this is an a", a bunch of times we gave it a bunch of examples of what an A looks like. And then we made a B above it and we said, "This is a B, this is a B, This is a B" and we taught it. We gathered data about what these things looked like. You can see here that we have x, y, and z coordinates for each of 20 positions in the human hand. And so you end up getting 60 features for one output and the outputs here are the letters of the English Alphabet A through Z. So I think y'all know what time it is; it's time to shout again. What kind of machine learning problem is this?
Yes, definitely supervised. I also heard someone say classification over there I forget who it was. But you did a great job. So yes this is a classification problem because there are a discrete number of things we're trying to predict. There's just 26 things and we know it's one of those 26 things.
So let's talk about how we went about solving this problem. The first thing we needed to do was choose a model and that's something that you'll often need to do when you're doing machine learning stuff. And we were somewhat early in our time of doing this kind of stuff so we weren't really sure what we were doing. We just took a bunch of models from scikit learn and just said try all of them and then we will evaluate all of them on the test data and pick the one that did the best. And that is actually, as I learned a little bit more, that's actually not a bad thing to try. Just try a bunch of different things and see what works best. Then once we did that it's not enough to just have a model; we had to build some sort of interesting application around it because it's not very cool to walk up to someone to say, "Hey, I made this thing that can tell you from handwriting data if you're making a sign for a certain English letter." That doesn't mean anything to anybody. And so we thought we would try to make a keyboard, and we were working on it and it would read out whether we were making an a, a b or a c or whatever and put those on the screen and we could see oh cool we're doing this.
It turned out that it wasn't quite accurate enough. And we even tried to do some stuff to do some Markov Chain stuff which basically takes into account the fact that certain letters are more common after other letters. So for instance "U" is way more common after "Q" than after a lot of other letters. Anyway we weren't able to get that to work very well and so we decided what if the answer here is to just try and try to make a different application around the same model. And we found that we could make a little educational game around this and we basically tried to market it as Rosetta Stone but for learning sign language. And this is a little demo that I will play. So it's a little game and it'll tell you to make a certain letter with your hand and then you try to make your hand look like that letter and then once you get it it gives you some points. At the end you can put your name in on a scoreboard because everybody likes to compete. Anyway, that's sort of what we built. And there were a lot of things we learned in this process.
The first thing that I didn't expect to be useful broadly because I thought oh we're just doing this for a hackathon. It's not a big deal. Limiting scope is huge. There are a lot of really huge problems in the world and if you try to tackle one you're going to just get lost. So it's really important to find some chunk of a problem that you can actually solve. Selecting a model is something that you're going to probably have to do if you're doing something like this. And this approach isn't the worst one. There's a lot worse you could do than to just try out a bunch of things and see what works best. The final thing I learned is that it's more than just the model. You can have a really interesting model that's good enough for a language learning game. And if you were trying to make a keyboard with it it wouldn't work very well. This is something that if you're working in a company you'll probably want to work with the product people in your company to decide what your users actually need, what they want, and what would be helpful to them. Try to gear your model toward that. Because at the end of the day we're all trying to make software for human beings.
Alright, let us move on to our second use case of the day. This is about forecasting energy load in the state of Texas. So the problem here is if we pretend that I am operating a power grid I have to know the demand at various places in order to be able to schedule the production of energy. We don't have excellent ways of storing energy for long periods of time so you kind of have to get things scheduled to where they'll be used shortly after they're created.
This isn't an entirely hypothetical problem. There is an organization that is known as the Energy Reliability Council of Texas and it is their job to manage the energy market in this state. For those of you who are here from out of town, Texas has (in most places) a deregulated energy market where power is generated and then sold on a market. And power companies buy it up and then they resell it to consumers. I'm not here to debate the advantages or disadvantages of regulation in the energy market. I think that would be a much longer talk if we were here for that. But the gist is that in a lot of places (Austin being a notable exception) the energy market is deregulated and you have to know when the when the demand is going to happen.
So ERCOT publishes a lot of data. They published the last at least 10 years of data I think it's 14 on energy load on an hourly basis in the various weather zones. So if we look at this, these colored regions are different weather zones because it turns out that weather is the biggest factor affecting energy usage because air conditioners are a wonderful blessing in my life but are also very expensive to operate. So they break this down on a weather zone by weather zone basis and on an hourly basis and then they also provide a sum (but I didn't care about that as much). And you could plot this on a graph and see over the last 14 years how much energy has been used on a daily basis in each of these different weather zones. It's kind of interesting honestly even at this point just looking at this graph I mean like, oh people are using more energy that's kind of an interesting thing to see and you can definitely see the seasonality of when summer rolls around in Texas. People use a lot more energy that's kinda interesting.
I think you are all prepared for what is next. What kind of machine learning problem is this? Regression! Yes, wonderful! This side of the room is killing it--y'all need to work on your game.
So a simple approach here to solve this regression problem is to just find the five nearest days to you and say that we'll take the average of those. And that's basically a k-nearest neighbors model where you find the five data points closest to a certain data point and average them together and then that is the output. And it turns out that because this is a time series we can set the data point value (basically the input) to be the number of day in the year it is. So for instance January 1st would be the first day of the year. Today is the hundredth day of the Year. Happy 100 days of 2018 everybody. And you can set that to be the input value and then the output value being whatever the energy load should be. This is sort of a time series data problem and being able to turn that into a regression problem is interesting. There's actually been a lot of study around that and this is a simple approach that is reasonable to go with (I think). When you're evaluating time series data there's a lot of things to consider. We obviously can still look at the error rates like a simple error rate just take our predictions versus what actually happened and see the absolute value of that and divide it by some reasonable-- like divided it by the actual number and then we can see we're off by 3 percent on average which is fine. I mean it depends on how accurate you're trying to get.
But another specific thing you'll want to do when you're working with time series data is look at what's called the residuals which is each of those individual data points where you take the predicted and the actual and you subtract them. And the goal is that there isn't a pattern in there if there's any sort of seasonality and that as you look at it you haven't quite fit the data as well as you could. The other thing you want to see is if your residuals resemble a normal distribution. If they're skewed one way or the other then you may have made a mistake somewhere. There's a lot of things I learned about this. The first thing I learned is that I did this in a very wrong way. You should really do a lot of research about this stuff beforehand. There's actually a lot of research on how to do stuff with time series data and the approach that I chose is actually not the most unreasonable way to do it but it isn't the best. There's a lot of really good tools out there like Facebook has a tool called Profit that is built specifically for predicting time series data and is used at Facebook we use it at my company and there's a lot of places that it's used and works really well. These libraries do a really good job of taking into account common things that happen in time series data. For instance holidays happen and energy usage is going to be way different on a holiday. So that's something to keep in mind.
The other thing I'll say is scaling the features is important. So for this problem my features were the number of day and the year that it was in were there were two input features and then I had my output. But the number of day it was is a different scale than the year. The year runs from 2014 to 2018 or whatever. And that's a different range of values than 1 to 366 potentially. That can cause problems when you're doing k-nearest neighbors stuff. Specifically, this is a similar example. This is actually the credit card application data we were looking at earlier. But just to demonstrate the problem when you don't scale your features. If we were trying to predict this yellow point on the left here you can see it far to the left and then a little bit above the red X. We're going to be looking at the features around us to try to find what point is closest to me that I can say that I'm going to be like that point. When we look at this with our human eyes we say obviously the red X underneath it is the closest point. But it turns out that the way that the features are scaled the net worth is such a larger number like just a bigger number than the age.
Age only runs up to 1 to 100 ish. Net worth can be a much larger range and so that has a huge impact on the distance. These numbers plotted next to each point on the left is the distance from the yellow point to the point associated there. So you see that actually the closest point is this one that's 50000 units away. And so it gets classified-- you can see on the right it gets classified as an "approve this application" even though it doesn't quite look like we should. What you can do though, there's a thing in scikit-learn called a standard scaler and it will take these things and scale them to to where the mean is zero and the standard deviation is one (which is really helpful in a lot of circumstances). So when you look at this visually they look the same because they kind of are; it's just the actual values have been scaled to a different range. And then when you look at the difference in how that ends up classifying, it looks more like what we would expect to happen. So scaling your features is an important thing to do (especially when you're doing something like k-nearest neighbors) but is also helpful when you're using other models.
All right. This is a recent project that I've been working on to use machine learning to find your next job. The problem that I ran into about a year ago was that I was passively job hunting.
Basically, I wasn't out there actively knocking on doors and handing out resumes or anything. And I was reasonably satisfied with where I was. But I was interested in hearing if there was a particularly excellent job out there that I might want better. I couldn't find something that did exactly what I wanted, and it seemed like I was getting a lot of noise coming through from just reading job listings. There were so many things that obviously I didn't want to look into. So I was wondering if I could make this a machine learning problem. I ended up doing was scraping a bunch of job listings. I would get the title and the company and then a link and then for a long time whenever I was bored I would just go to this spreadsheet on my phone, click on the link, read the job description and then come back and say whether or not I thought it sounded cool. I gathered a bunch of data like this and if I'm being honest with you I probably spent more time reading job descriptions this way than I would have if I didn't build this. But because all of you are here I don't think any of you have room to talk about over engineering something. So if I can't talk about my love for this here I don't think there's any safe place for me.
Anyway, you should be familiar with this question by now, what kind of machine learning problem is this?
Classification! Yes, wonderful. Thank y'all. I heard someone say clustering. It's not quite clustering because we do have a specific output variable that we're trying to predict. We want to know whether a given job sounds cool or does not.
The way that I ended up solving this is kind of tricky. So if we look at the other problems we've seen today they're all numerical data, right? We have a day number and a year. Those are obviously numbers. We have a net worth and a loan size; those are numbers. Age is a number; net worth is a number. How do we turn a job title into a number? Computers can't deal with text. You can't just throw text at a computer and have it know what to do. You have to find some way to turn it into a number. And when I ran into this I was thinking what am I going to do here. I have all this data about text but I don't know how to fix that. So I turned to our trusty friend Google and I searched "text representations for machine learning" which pretty much is the way that you should learn a lot of stuff is just search search it up. And I found this. This is an idea that is a good thing to try first. It's not state of the art by any means, but it's a good first pass. It's called a word count vector, or people will call it a bag of words.
And basically what you do here is you take all of your job applications, you find every word that occurs in any of the job applications, and place those along the columns of a matrix. And then you place each job title along the rows of the matrix, and then to fill in each slot in the matrix you look at the job title and the column and say, "Does the word in the column appear in the job title on the row?" So for instance this first one: engineer does not occur in that title but web does and applications does and senior does. And I'm not going to walk you through filling out this matrix because even after 4 I'm a little bit bored of it. And one thing to note though is that while in job titles usually they don't repeat words you theoretically could. So I put that last one I don't know who's going to post a job titled "Data Data Data Data," but I'd be interested to hear about it. And so you can see where there' multiple occurrences of a word it isn't just a one it can be you know four or whatever. So. That's basically how we can turn the text and the boolean value into numbers. So that's this highlighted green part it becomes this series of numbers here and the highlighted blue part becomes a number there. And then it's really surprisingly simple to do this stuff because a lot of this functionality has already been built for us because there are giants upon whose shoulders we can stand and see much further.
So this is a really simple example where we take our rated jobs, pull out the titles, and then pull out whether or not it sounded cool, and then scikit-learn has this tool called a count vector which will take text data and turn it into those word count vectors I was talking about. And then we can take that data and put it into a model and fit it with the preexisting "sounds cool" or "not" data. All we have to do then is just predict on the data, and we get out this array that I've highlighted at the bottom that says, "OK the first job in the list you gave me doesn't sound interesting but the fourth one does." So that's the model I ended up building. And originally I was just doing this in a Jupyter notebook, and I was just running through it and I got a classification error of 19 you know around 20 percent was like heck yes, I'm God's gift to data science. This is going to be amazing. And then what I realized was it was just saying that everything didn't sound cool. And what I realized is most of the job listings I was reading didn't actually sound that interesting. And so I would just rate them as they didn't sound cool and the model picked up on that and it's like, "I can do super well if I just say nothing sounds cool."
So I committed what's called the base rate fallacy. And this is something that's important to understand when you're approaching a problem like this is to understand what would happen-- what are the underlying rates in these problems. Because I wasn't actually improving on anything. I was just doing as well as literally just guessing zero every time. So this is a self-portrait I drew after I discovered that I made this problem. Dealing with imbalanced classes like this is fairly common. And so I wanted to provide a little bit of insight into good ways that people do this and the way that I ended up doing this. The first thing that you can do is use better error metrics. The only way I realized that I was having this problem is because I knew to look for the base rate (and now all of you know to do that). But these are metrics that will help you understand your data in a little different way. There's no one metric that's going to be perfect for every situation, but having a family of them can help you understand what's going on much better.
Precision and recall are related concepts. In our case precision means how many of the job titles that I said sounded cool actually are cool and recall means of all the job titles that do sound cool. How many am I saying sound cool. These give you a better understanding of how you're doing in terms of like false negatives and false positives and stuff.
The other thing that is useful is to use what's called the confusion matrix which you can see at the bottom here and what you do there is you put the predicted values on one axis and the actual values on the other axis. And if I were to do something like this I might see well I'm predicting zero-- and, actually, I filled this out wrong-- but I'm predicting zero for everything, and I would notice that error much more quickly.
Other than using better error metrics, one thing you can do is called "under sampling" and this is what I actually ended up doing. I had 500 job titles (let's say) and only 100 of them sounded cool. I took all of those and then I took 100 randomly selected not cool job postings and I made a new dataset out of just those 200 and I trained a model on that and that got me a much better accuracy rate for the jobs that did sound cool. Another technique that people do use is called oversampling which is kind of the opposite. So if I have those 100 cool job postings I would take those and have four copies of each of them, so that would give me 400 cool postings and then 400 not cool postings and I could just train my model on all of that. I've never actually used that because I feel weird about doing that but it's something you can do if you want to.
So in the end what I ended up doing was getting this all running in essentially a cron job on a remote computer and it will every week email me just a list of the top 10 jobs that sound the most interesting. So this is another thing where we talk about how do we want to use this model. If I were to just have it spam me all the jobs it sounded cool that would be more than I want to look at. But, because I chose a model that can predict the probability of something. Logistic regression is able to tell you how probable it is that a certain job sounds cool, and I could just pick the 10 that sounded that had the highest probability of sounding cool. That gives me a much shorter list to look over each week.
So some lessons that I learned from this. Obviously the first one is understand the base rate because that can really make you sad. The second thing is that doing something simple doesn't mean that it's going to be ineffective. Do any of you watch The Office, or have any of you watched the Office? OK, so there's a scene in there where Dwight is talking about Michael and Michael is his boss and his boss comes to him and he says, "K I S S keep it simple stupid. It's great advice and it hurts my feelings every time."
And that's kind of how I felt about it. I'm like I want to do something cool, I want to do deep learning, man! But it ended up just being good enough, and using a very simple model worked.
So the approximation generalization tradeoff is a theoretical concept from machine learning that can help us understand why this works. And as you might guess from the name it means that if you have more approximation you're going to have less generalization; if you have more generalization you're going to have less approximation. Those words don't really mean anything so I drew a graph that will help us understand. Again, here I've made up some data. The blue line is the truth. And then I sampled some points from it with a little bit of random noise in there to again model the real world. What I did was I fit two different models to it. One of them is a simple model, linear regression, which you probably learned in a high school algebra or precalc class. And then one of them being a more complicated model which can effectively memorize any data set that it wants to. What you see here is that for the points that I'm showing right now, the red model is killin' it. It knows every single spot; it has zero error on those. It is approximating the data set extremely well. It knows the training data by heart.
However, what you may notice is that when we add more data to this it does not do as well on those points. What you see
is the green model doesn't do as well on the input data as the red model does, but it does much better on the out of
sample data (on the testing data). So we have this tradeoff where more simple models are generally better at
generalizing even though they're worse at approximating. So that's sort of why it's a good idea to start out with
something really simple and basic and work up from there. The other good reason to do this is that it's easier.
scikit-learn is a
conda install away versus, I mean in my experience, setting up TensorFlow is hard. And even once you
get it set up training stuff can be sad and hard and long, and it's just a lot easier to start out with something simple
that you can iterate on quickly and learn and learn a lot about your problem space before you go into something more
So we've just talked about a good amount of things. This is sort of a summary that we can view some key concepts from each of these use cases we've talked about. For the teaching a computer sign language, what we ended up doing was support vector machines (which is a model that is useful). It's built into scikit-learn. In the forecasting energy load in Texas data, it was time series data and what we found was using k-nearest neighbors worked really well. code:2:4
However if you're doing time series data you should probably do some more research and probably use something like Profit that's specifically built for time series data. Then the last use case we just talked about. If you run into text data, it's at least worth trying Bag of Words. It has its caveats; it has its downsides, but it's a good first step. And I ended up using logistic regression and that works really well and I get the email every week and I'm happy with it. So it works pretty well.
So basically there are some takeaways I have here. And then some recommended tools that we'll talk through. The big takeaways being (from the very beginning) in supervised learning, we want to use past examples to predict a continuous value in the case of regression or a discrete value in the case of classification. And those two correspond with questions like "how much of this thing?" or "what kind is this thing?". And then another huge takeaway is to try the simplest thing that could possibly work. This is something that my machine learning professor tried to beat into our heads and has proven to be very effective in my experience. Once you have that simple thing that is kind of working you can always test it out and iterate and maybe try a different model maybe try a different set of features and work from there.
We've been kind of light on recommendations about specific tooling but just if you want a jumping off point Jupyter notebook is a great tool that lets you interactively run models and train them on various datasets and see how they look kind of. There are some some plotting tools like matplotlib and Bokeh that will let you see into what the data sort of looks like and can really help you get a better intuitive understanding for what's happening under the hood. Pandas is a great library for manipulating tabular data which, actually, all the data we saw today was all tabular data in that it had a set of rows and a set of columns. Pandas does a really good job of handling that kind of data. It can do things like read from Excel spreadsheets and read from HTML tables and read from CSVs and whatnot. Obviously I recommend scikit-learn. I used it for all of this stuff, and it's nice to not have to reimplement this stuff yourself.
There's a lot more resources available if you're interested in this stuff. If you're interested in more of the theoretical side I highly recommend a book called Learning from Data. It does a really good job of treating machine learning theory with respect. A lot of times when we talk about machine learning it feels like we're just pulling stuff out of a bag or pulling out a bag of tricks, and it's not really fair to think about it that way and there's a lot more to it than that. This does a good job of helping you understand how that works.
On the opposite side there is a blog called Practical Business Python that talks a lot about how to use these specific tools and if you're hungering for more after this talk about how to specifically do stuff. He has a lot of great resources about "how do I graph something? How do I read an Excel file?". It's really interesting, really good, solid, extremely practical, detailed read there. Then the biggest thing I would say as far as gaining extra experience from others is reading the Kaggle blog. They call it no free hunch (which is an adorable name) and they have a specific section on it for winners interviews which is where-- there's all these people who compete in machine learning competitions and then whoever wins they'll do an interview of them and say what did you do. Reading through those is a huge amazing resource that I don't think is being taken advantage of enough because you can learn from some of the best data scientists in the world about how they do their job and then apply those in your own work. If you're interested to hear a little bit more detail on the sign language or machine learning to find your next job part, these are links, and I'll tweet out the slides in just a little bit.
I have more information on my website, samueltaylor.org, if you're curious about those things. I'm also happy to talk to you afterward if you have any things you want to talk about. I do work for Indeed, and I would be remiss not to thank them for their support of me doing this kind of work and talking about this stuff in front of people. If you are looking for a job please come talk to me. We like data stuff. Beyond that, again I'm Samuel Taylor. I prefer communicating over email over pretty much anything else so if you have a question you're obviously welcome to come talk to me right now but if maybe you're a little shy or just don't want to talk feel free to email me. I love reading email. I might be the only person who loves email like that. And then also I am happy to read people's tweets-- if you have questions I'm happy to take those via Twitter as well. I'm @SamuelDataT. Would love to hear from you.
Thank you so much for letting me talk to you and take this time out of your day. I appreciate it so much. I really hope you're able to get something out of this if you have any other questions. We have about 5 minutes that I can take questions if you have them or I'm also happy to talk about it after this, but thank you.