(If mathematical formulas are not showing up here, please read this on the web. .) Here's a scenario that I believe to be common. I've got a dataset I've been collecting over time, with features \(x_1, \ldots, x_m\) This dataset will generally represent decisions I want to make at a certain time. This data is not a timeseries, it's just data I happen to have collected over time. I will call this…
(If mathematical formulas are not showing up here, please read this on the web. .) In a previous job, I built a machine learning system to detect financial fraud. Fraud was a big problem at the time - for simplicity of having nice round numbers, suppose 10% of attempted transactions were fraudulent. My machine learning system worked great - as a further set of made-up round numbers, lets describe…
(If mathematical formulas are not showing up here, please read this on the web. .) It's become a popular meme that "shareholders only care about the next quarter". Lots of people make arguments like this - for example, Jamie Dimon and Warren Buffet . As the meme goes, shareholders only care about the next quarter of earnings, and CEOs make decisions accordingly - sacrificing long term…
(If mathematical formulas are not showing up here, please read this on the web. .) A lot of suspicious behavior can be detected simply by looking at a histogram. Here's a nice example. There's a paper Distributions of p-values smaller than .05 in Psychology: What is going on? which attempts to characterize the level of data manipulation performed in academic psychology. Now under normal…
(If mathematical formulas are not showing up here, please read this on the web. .) Frequently in data science, we have a relationship between X and y where (probabilistically) y increases as X does. The relationship is often not linear, but rather reflects something more complex. Here's an example of a relationship like this: In this plot of synthetic we have a non-linear but increasing…
(If mathematical formulas are not showing up here, please read this on the web. .) Lambda School has recently come under fire by the mainstream media for having success rates smaller than 100%, as well as for having a founder who is a nerd. The articles imply that Lambda School is somehow ripping off it's students - possibly by, um, tricking hedge funds into paying for their education - though no…
(If mathematical formulas are not showing up here, please read this on the web. .) I have recently been working on setting up a trading strategy and running it in the cloud. Although I haven't used Azure before, I wanted to try it out - some of the data science features that Microsoft advertises look pretty nice. This post is not of general interest, and most readers should stop reading now. This…
(If mathematical formulas are not showing up here, please read this on the web. .) I was recently discussing a project with a younger data scientist and I noticed a curious mismatch between our language. We had an API that we wanted to impose rate limits on. We want to ensure that 99% of our good customers have a good experience and never hit the rate limit, while locking down the heaviest users…
(If mathematical formulas are not showing up here, please read this on the web. .) I recently gave a talk at the Fifth Elephant 2019 . The talk was a discussion about how to use math to handle unfixably bad data. The slides are available here. . Go check it out.
(If mathematical formulas are not showing up here, please read this on the web. .) With Andrew Yang's presidential candidacy moving forward, people are discussing basic income again. One common meme about a Basic Income is that by removing the implicit high marginal tax rates that arise from the withdrawal of welfare benefits, disincentives for labor would be reduced and therefore a Basic Income…
(If mathematical formulas are not showing up here, please read this on the web. .) I recently saw a silly twitter exchange between two of the lyingest politicians in American politics. Given that they have both explicitly expressed the viewpoint that morals matter more than numbers and being "technically correct", I figured that I should just check for myself. On twitter, Trump says white…
(If mathematical formulas are not showing up here, please read this on the web. .) Anyone following Nassim Taleb's dissembling on IQ lately has likely stumbled across an argument, originally created by Cosma Shalizi , which purports to show that the psychometric g is a statistical myth. But I realized that based on this argument, not only is psychometrics a deeply flawed science, but so is…
(If mathematical formulas are not showing up here, please read this on the web. .) I recently gave a talk at CrunchConf 2018 . The talk was a about the various impossibility theorems that make a person concerned with AI Ethics must content with. The slides are available here. . Go check it out.
(If mathematical formulas are not showing up here, please read this on the web. .) I recently attended a discussion at Fifth Elephant on privacy. During the panel, one of the panelists asked the audience: "how many of you are concerned about your privacy online, and take steps to protect it?" At this point, most of the hands in the panel shot up. After that, I decided to ask the naughty question:…
(If mathematical formulas are not showing up here, please read this on the web. .) I recently gave a talk at the Fifth Elephant 2018 . The talk was an introduction to linear regression and generalized linear models from the Bayesian perspective. The slides are available here. . Go check it out.
(If mathematical formulas are not showing up here, please read this on the web. .) In principle A/B testing is really simple. To do it you need to define two separate user experiences, and then randomly allocate users between them: def final_experience(user): if random.choice([0,1]) == 0: return user_experience_A(user) else: return user_experience_B(user) So far this seems pretty simple. But then…
(If mathematical formulas are not showing up here, please read this on the web. .) Today I spoke about AI ethics at 50p 2018 . Here are the slides from my talk. The general topic was multiple ethical principles, and how it's mathematically impossible to satisfy all of them.
(If mathematical formulas are not showing up here, please read this on the web. .) Some time back, I was involved in a discussion with folks at an India-based software company. An important question was asked - why isn't this company as productive (defined as revenue/employee) as it's western competitors, and what can be done to change this situation? In this discussion, I put forward an…
(If mathematical formulas are not showing up here, please read this on the web. .) I recently analyzed a somewhat puzzling data set. I was sending HTTP POST requests to a system. The system's would then acknowledge receipt of these requests (returning a 200 status code), and some time later (it was a slow asynchronous process) send a web hook to a specified URL if the request was successful .…