This is what I came away with after working on the Netflix dataset.
Incentives, not perceptions, are what need to be changed for realizing better recommendation engines (or collaborative filtering algorithms; whatever you call them).
Negative recommendations are not a good idea. Unlike false positives, the consumer really has no way of knowing whether a negative was in reality true or false. Recommendations are also about feeding our ego , even if the engine is ultimately basing its suggestions on consumers whose tastes are very similar to yours.
It is indeed magical; consumers who think they are expanding horizons by seeking out indies or documentaries on Netflix are simply discovering latent interests that have already been explored by others in your interest cluster.
The author is right though; false positives are what weigh most heavily on our perceptive scales. Risk aversion ensures that people do not venture out beyond what are suggested to them (by an automated agent or friend), and collaborative filtering algorithms have a cold start problem. For a movie to appear on the radar, enough people need to have watched and expressed enjoyment. Digital filtering makes it even harder for random exploration that is needed to seed these. Failing that, you get middling recommendations like that of Netflix, where the average rating (~=3.6) is no better that what you would get without it.
There is one easy way to get around the risk aversion obstacle. Offer "free" movies from a different, high variance list of movies, with "free" meaning a fourth movie on a three-per-month subscription, and variance being high on the predicted ratings of the consumer. Such an offer, while being marginally more expensive for Netflix, will allow consumers to experiment without any perceived costs (which, currently is the wait for the next Wire or Deadwood DVD you could have gotten otherwise).
If I may speculate, recommendations should be about gentle, guided exploration and not avoidance.
I wonder if an expert (human) with access to the same data set, any automated tools they choose and their own knowledge/research of movies could do a lot better. I think they probably would be able to crack the 10%. If they couldn't then it would be a fair indication that the data is simply too noisy to beat the 10%.
Interesting... even though accuracy is generally high, the false positives weigh heavily on the experience.
How accurate can we get recommendation systems anyways? I know my friends recommend movies to me all the time and a lot of them aren't my taste - do we expect computers to outperform our close acquaintances?
Incentives, not perceptions, are what need to be changed for realizing better recommendation engines (or collaborative filtering algorithms; whatever you call them). Negative recommendations are not a good idea. Unlike false positives, the consumer really has no way of knowing whether a negative was in reality true or false. Recommendations are also about feeding our ego , even if the engine is ultimately basing its suggestions on consumers whose tastes are very similar to yours. It is indeed magical; consumers who think they are expanding horizons by seeking out indies or documentaries on Netflix are simply discovering latent interests that have already been explored by others in your interest cluster. The author is right though; false positives are what weigh most heavily on our perceptive scales. Risk aversion ensures that people do not venture out beyond what are suggested to them (by an automated agent or friend), and collaborative filtering algorithms have a cold start problem. For a movie to appear on the radar, enough people need to have watched and expressed enjoyment. Digital filtering makes it even harder for random exploration that is needed to seed these. Failing that, you get middling recommendations like that of Netflix, where the average rating (~=3.6) is no better that what you would get without it. There is one easy way to get around the risk aversion obstacle. Offer "free" movies from a different, high variance list of movies, with "free" meaning a fourth movie on a three-per-month subscription, and variance being high on the predicted ratings of the consumer. Such an offer, while being marginally more expensive for Netflix, will allow consumers to experiment without any perceived costs (which, currently is the wait for the next Wire or Deadwood DVD you could have gotten otherwise). If I may speculate, recommendations should be about gentle, guided exploration and not avoidance.