At my day job (at matchmine), we’re working on creating a better way to help people find the stuff they like. Sure, it’s infinitely more complicated than that, but I think that’s a pretty good single sentence description. Because of that, I’ve been spending a lot of time trying to understand the problems involved in helping people find stuff they like.
Time for an ENORMOUS disclaimer:
In the following post, I’m coming at the preference question from the perspective of a guy that has absolutely zero understanding of the statistical side of recommendations/preferences (which, I’m told, is the single most powerful way of delivering recommendations). Half of what the science guys say to me goes over my head. The other half goes WAY over my head.
When I showed this post to the science geniuses here, I was told that I should really add a note about the difference between understanding and statistical probabilities. I was told that within a system with a large number of content items, it could make sense to "abdicate understanding in favor of statistical probabilities of connections." While the statistical solution may actually work and deliver on the promise of great recommendations, the statistical model does not address the "why" question.
So with that said, I’m kind of stumbling around in the dark trying to make sense of the things I like. I’m starting with my own experiences as a consumer, and trying to back in to an understanding of how recommendation systems would work as a human being. Basically I’m ignoring the highly complex mathematics, and thinking of discovery as if I were trying to recommend media items to a friend.
1. You have to understand what people like- If you’re trying to find stuff someone is going to like, you have to understand the kinds of things they currently like. For instance, if you want to recommend a movie to me, you have to know what kinds of movies I already like, right?
2. You have to understand why people like the things they like- If I say I loved The Big Lebowski, that would certainly help you in coming up with other movies I may enjoy. But you really need to know why I liked it. Is it because it’s a Cohen brothers movie? Because I’m a huge fan of The Dude? Am I really into bowling? Or am I sucker for any movie with an amphibious rodent? To give me a good recommendation, you need to know what I like as well as why I like it.
3. You have to understand what people do not like- Knowing what someone does not like can be a huge factor in understanding the things they will enjoy. For instance, I really like comedy, but hate sitcoms. I actually become enraged when I see someone watching "Two and a half men." The fact that people can watch a show that worthless actually disgusts me physically. So knowing that I love comedy but hate formulaic, laugh-track laden drivel would certainly help filter out those things that are considered "comedy" but make me angry.
4. You have to understand why people dislike the things they dislike- Why do I like stand-up comedy, but hate "Two and a half men?" Like I said above, they’re likely to both be classified in the "comedy" bucket, but one makes me laugh and the other makes me violently angry. Where some recommendation systems would simply offer categorization-based suggestions, the only way to truly help someone "discover" content they will love is by really understanding the difference in preferences between seemingly similar items.
5. You have to understand the connections between liked items- To truly deliver on the promise of "discovery", a system needs to have a deep understanding of how liked items are linked together. A simple categorization or genre rating will not do. Here’s a hypothetical- My two favorite songs are Alice’s Restaurant and Boy Named Sue. The former would likely be considered Folk, while the latter would be seen as some type of country. On the surface, the two songs would seem very different, but if you look at each song’s attributes, you’ll start to see the connection- both are long, humorous songs told in story form, and performed live.
6. You have to understand the connections between disliked items-Just like understanding the links between two liked items, understanding the shared attributes of disliked items can help filter out what would seem to be enjoyable recommendations. I personally hate all Madonna songs. All of them. Seriously. I also have a white-hot hatred for U2. Why? They seem unrelated, right? Maybe I hate them both because they are too dramatic for me. I hate the way Madonna over-annunciates, and loathe how Bono stretches out every word as if he’s paid by the syllable. Perhaps I dislike both because their lyrics are absolutely predictable in every way, filled with obvious cliches at every turn. Yep. That’s why.
7. You have to understand the connections between liked and disliked items- I think that we humans are inherently contradictory. I might hate American Idol-like singers, but if I hear my friend’s cover band (Red Square) playing Carrie Underwood’s "Right Now", I absolutely love it. Same song, different performer. I may claim to hate every hippy jam band, but really enjoy when String Cheese Incident covers a rap song.
8. You have to understand the "preference decision threshold"- I am at a loss for a better term for this one, but here’s the basic premise: If I’m given a song to listen to, I may not love it. I may not hate it. My feeling about the song may be lukewarm. Does that mean I like it? How about a movie? If I love the first hour of a movie but hate the ending, does that mean I like the movie? Dislike it? What is my threshold for deciding whether I like or dislike an item that I’m not passionate about?
9. You have to understand transient liked items- There are some things I like today that I’ll dislike tomorrow. The example that comes to mind is breakup music. Let’s say I just got dumped. I might want to sulk and listen to Dashboard Confessional, but a week from now, I’ll be in a great mood and hate that kind of music. We like different things at different times, and when there are items that are closely associated with a mood, making recommendations based on them can be hit or miss.
10. You have to understand the social value of liked items- This last point could be an entire book on its own. It involves whether an item is liked on its own merit, or whether the item is liked based on what liking it says about you socially. Record sales show that lots of people liked the Backstreet Boys or the Spice Girls. But how many people would admit to liking them? This is the notion of a guilty pleasure- liking something behind closed doors, but never admitting to it.
These are just some examples of things that need to be considered when attempting to understand what people will want. Again, this post is a complete oversimplification of trying to understand what people like. I’m certain that I’m missing important points (including the gaping hole that is anything to do with math and statistics) and I’d love to hear any comments or questions.