When I was about 13 years old my father took me on a holiday to Singapore. I had a blast; it was a new country with new sights, sounds and experiences. I also learned an important lesson.
If you haven’t been to Singapore, one of the (many) cool things about the place is the food. And so, when in Singapore, you eat. A lot. And at one restaurant we were presented with a rather interesting dish. It was brown, goopy and spongey, and my thirteen year old brain was like: “that’s gross!” I then verbalised this sentiment.
My dad’s response: “I don’t care if you don’t like it as long as you try it first.”
Me: “I don’t like it.”
Dad: “You haven’t tried it yet. Try it, then tell me you don’t like it.”
I ate it. I didn’t like it. But the lesson stuck.
Fast forward about 15 years. Personalised content and custom recommendations are commonplace. Whether it’s Netflix making recommendations or Pandora allowing you to tailor your listening preferences, you’re watching, reading and listening to the things that you know you’re likely to enjoy before you try them. Sounds great. But what about trying different things?
I’m worried about a future where everything is tailored to your taste. Tastes and preferences can and should change. For example, the music I listened to at age 16 is different to the music I listen to now. And no – it’s not just that my taste is maturing due to age. It has a lot to do with a desire for exploration, receiving recommendations by friends, and by generally taking joy in exploring things outside my comfort zone.
Not all of us venture out of our comfort zone, and the motivations for doing so are pretty interesting in themselves. But for me – I love discovering new things. There is a thrill in tasting something new, seeing something different. Many of my friends and family feel the same joy in discovery.
So what happens when we’re systematically pushed towards a new choice based on past choices? Ah, yes… you liked Blah so you’ll like Bleh, right? Rather than growing and changing as people, we become static. Your tastes solidify and you become just like everyone else in a particular category. We become our own little echo chambers, with every choice informing the next choice, waltzing down a path of algorithmic satisfaction and zero adventure.
I’m not trying to make a scientific argument about recommendation engines, or say that personalisation is bad, or that we shouldn’t enjoy the recommendations we receive from algorithms. I just want to make two points.
First, I want an anti-recommendation engine. Something that shows me something I probably WOULDN’T have tried. An algorithm that says, “Based on your past choices, here’s something you haven’t tried and might not like but goddammit you should try it anyway.”
Second: Don’t knock it till you try it.
You could argue that a good recommendation engine is one that presents you with stuff that you should try… whether that stuff is similar to your previous selections, or something completely different. It shouldn’t overweight based on similarity.
I agree – and I think that’s what i’m somewhat flippantly asking for. The problem is that (modern) recommendations generally rely on your behaviour and ratings and on ratings of people similar to you, and so the algorithm needs to produce a recommendation based on that data. Isn’t homogenization inevitable, at least within some kind of local “taste group”? Unless you and people like you already display random behaviour, and have explicitly rated the edge selections highly, you’re not going to get anything completely different, but things that “fit” your tastes.
Ultimately though, the argument is more theoretical than practical at the moment: personal recommendations are the ones that generally have pushed me outside of my comfort zone, and they aren’t going anywhere… but if you imagine a future where we are solely relying on algorithms and computers, you’d need to train some of the machines to produce challenging recommendations to achieve the balance you mention.
So, Ryan has a great point. A great recommendation engine will recommend things that you will like. (That are new to you). So it should encompass things that you haven’t tried but be able to predict things you will like. Granted, it can only go off data about you, or about the crowd. If it suggests something to you 1 hour after you just survive a heart attack, induced by your first mind bending LSD experience at age 40 that you took because you were having a mid-life crisis – then it will probably suggest the wrong content. (Unless the prediction engine is reeeeallly good – maybe it knew your search terms.)
But there is still plenty of good adverture in this, even if the path has been trodden before. It could even recommend stuff that nobody has seen by assessing artists you would like that just released stuff to YouTube etc.
I think the real fear here is the loss of serendipity – despite this, you still have it in the new world.
I recently got an email from an old friend that came across my blog when searching for a Fitness First ad. Digital Serendipity.
Then there is the loss of the path less trodden. Sure, going to Machu Picchu isn’t as adventurous as it used to be (or Everest), but when it is your first time – it is still an adventure.
Fortunately, while the real world is becoming more trodden, the digital world is expanding at a rate that is less trodden. This is a great thing, because it means your next digital adventure through technology and culture starts right here. Next stop, http://www.damiendonnelly.com
But again, i’m not so much talking about recommendation engines not turning up something new or something you haven’t watched. Ideally the engine knows what you’ve seen and will always give you new options.
I’m talking about the fact that a recommendation engine gradually gets to know what you like, and over time learns your “sweet spot”. I’m arguing that this is not a good thing, because while you still may be watching new movies or hearing new songs, you’ll gradually home in on specific categories that it’s known you enjoy, and so you won’t be offered very different types of content.
Consider pop music. Plenty of songs have the same chord progressions as other songs. This is because they’re known to sound good to the general public. Many, many hits have shared the I-V-vi-IV progression. This isn’t a bad thing. But it’s the reason so many pop songs sound the same, and it’s the reason people say you only need to learn 4 chords to learn the guitar. Record execs have learned that songs based around that structure are more likely to be hits. For that reason, last summer the same producer produced two hit songs – one for Kesha, one for Katy Perry, and both are onefivesixfour-ers. He’s completely playing to people’s most basic tastes to give his client a hit.
Now imagine that the music producer is the recommendation engine. You are the general public. Because the engine has learned what you like over time, it will give you different choices, but with similar underlying structures because it knows your tastes and wants to have a “hit”, at least as far as you are concerned.
That’s the problem that I see, and unless you explicitly train your recommendation engine to provide some slightly adversarial results, our tastes become homogenized within known taste categories. Which sorta sucks.
Another way to say it is that the customer doesn’t always know what they want, and we should try to train recommendation engines to take that into account.
Nice site Brett :)
Sooo… By virtue of inaccuracies that exist in ANY and EVERY system, there should be some natural level of irregularity. For example, I order a Scotch and dry at 4am during a big night out, however, the bartender presents me with a Scotch and milk! Now, I would like my Scotch and dry but I am compelled to try said Scotch and milk even though I know it’ll taste like pure concentrated evil…
So to recap, I know I ordered the correct drink (common sense may say otherwise due to the obvious, but for the sake of the exercise lets run with it =P ) but somewhere in the system this message became muddled, resulting in the liquid horror before me….
This could occur in any system that is reliant on human input somewhere in the process, whether it be in the system definition or programming (for you code monkeys =P ), or from the end user’s incorrect description or tag. Irregularities exist in even the most stringently designed, checked and verified, tested, checked and verified, commissioned, checked and verified, and verified and verified, and maybe checked and verified again of implementations. And this is in comparrison to a recommendation engine that I somehow doubt is mission critical!
So fear not Brett, human error will save us! :)
[...] http://brettwelch.com/2011/01/13/recommendation-engines-and-the-death-of-adventure/ Rate this: Share this:EmailTwitterFacebookMoreStumbleUponRedditDiggLike this:LikeBe the first to like this post. [...]