Right now, I should probably be writing that promised article on what makes a good advertisement, but instead let’s talk about the magic of Machine Learning and people.
We’ve all heard the hip, new cool thing is machine learning. It’s PyTorch or Keras or whatever other flavor that’s out there. I got taken up with it, too. Heck, I even have a certificate around here that says I survived learning how to use Tensorflow. All this means is that I’ve peered behind the curtain a bit to get a glimpse of the little man at the controls.
It is a wonderful technology, to be sure. It allows a great many conveniences that will or have already lead to some pretty amazing insights. Because what it does is essentially multivariate regression analysis – but at a massive scale. That means it basically line-fits large sets of data so that it can determine the statistical probability of similarity to other things – not just the one or two variables Excel can do.
I’m sure right now you’re thinking that was an un-welcomed digression that’s left an accounting term-like mental malaise in its wake. What’s the point? Well the point is this: While the tech is good at boiling down complex data sets to provide you with, say movie choices that you may like based on what you’ve already liked, I don’t think anyone has factored in people changing their tastes from time to time – or the fact that while there may be a large number of people looking at things, there’s only a small amount of things they can look at.
Examples could found in a variety of places, perhaps the most memorable for all of us digitally plugged in would be either online dating, shopping recommendations from online retailers or streaming services like Netflix. By now, I’m betting you have had an experience with at least one of these. That means we’ve all experienced the pain of only getting exactly the things we’ve liked in the past. The trouble is what I liked to watch five years ago may not be what I want to watch now. Or, it has finally dawned on me that the type of person I was into back then is not the best choice going forward, but technology has doomed me to stay in the same dating pool.
When will these machine learning things figure that out? I think we all feel like they never will.
Why? Because the very large data sets that give machine learning tech their hyper-focused funneling power is the very same massive data set that truly has to turn over before our new tastes are included – and that’s only if the programmers of these systems have put in place code that allows for this to happen.
If you look at the machine learning math (and I won’t open that Pandora’s Box), it’s easy to see watching only one Tom Segura stand up special has no chance of offsetting those ten Jeff Foxworthy shows your uncle watched on your account. Of course, that’s IF they built the software to recheck your preferences from time to time.
To really simple down the math (and yes the math is much more complex than this, Mr. Data Scientist, but it’s an analogy for everyone else), the effects are similar to every time Foxworthy spends 30min explaining what a redneck is on your Netflix account the software multiplies the Foxworthy counter by ten. The same would happen for watching Tom. So in the easy math world it looks like this:
One viewing of Jeff: Jeff * 10 = 10Jeff
One viewing of Tom: Tom * 10 = 10Tom
Things get bad when that scales – as in your uncle just watched 10 Foxworthy specials on your Netflix account:
10 viewings of Jeff = 10Jeff^10 =10000000000Jeff
Compare this to one 10Tom and one viewing of something else isn’t going to change what the machine learning system is thinking you like, but what more important is that there isn’t 10 Tom Segura specials, there’s only like five.
Five Tom specials = 10Tom^5 = 100000Tom
10000000000 > 100000
The exponential math here means that there’s really no way watching less than the same amount of something else is going to offset the binge of a previous show, genre, actor or however else they’ve segmented their content.
Sorry, you’re doomed to finding out (over and over again) refrigerators on your porch probably makes you a redneck -unless Segura gets off his ass and makes five more specials.
Of course we’re all really in for it if they’ve also added in counting the frequency between views of the same materials. If that’s the case, binge watching is particularly dangerous.
All of this content railroading is compounded by these services continuing to only show you selections that fit with what you liked previously. There seems not to be any sort of random content display to break you out of your filter bubble. This makes you have to independently locate different selections – which is hard because how would you search for something when you don’t know what it is?
I’m hoping that eventually the powers that be will figure this out, because, while I really like Korean action movies (and you should check them out, too), I’d like to have the opportunity to see Thor: Ragnarok without having to spell all that out with the ridiculous Roku remote every time.
Eric Fescenmeyer's Blog 