Pretentious enough title for you?
I
Recommender systems – those algorithms that guess what you may be interested in as you browse Amazon or listen to last.fm – are commercially important. Netflix claims that 60% of its rentals are driven by its Cinematch recommender system [link]; that’s over half a billion dollars of business in 2008. As online commerce continues to grow, recommender systems will only get more important.
Recommender systems are culturally important too. As more of our culture moves online, they will be responsible for more of our cultural experiences, and will play an important role in shaping the creative parts of our societies.
Recommender systems will get better. Ten years ago they were largely improvised. Now you can do a Ph.D. in recommender systems and there are international academic conferences all about them [link]. The subject is ideal for academics – it is algorithmic and yet open ended, with many different approaches and criteria for success. It’s an endless playground for exploration and simulation.
Even though they will improve, there is no such thing as an optimal recommender system. Accuracy is insufficient. The interests of recommendees vary. Serendipity, intra-list variety, reliability and trust-generation are just a few other considerations [pdf link].
Don’t confuse the outcome of recommender systems with intrinsic merit. The recommendations are highly dependent on history and are the products of cumulative advantage. Many think that “if the experts could only figure out what it was about, say, the music, songwriting and packaging of Norah Jones that appealed to so many fans, they ought to be able to replicate it at will”. But hits cannot be reliably predicted because our choices and preferences are too inter-linked. Clive Thompson writes that companies with recommender systems can “track everything their customers do. Every page you visit, every purchase you make, every item you rate — it is all recorded.” [link] But other studies have shown the systems to be chaotic. Tiny, random fluctuations can lead to completely different outcomes. [link]
Recommender systems can easily reinforce inequalities among recommended items. A system that recommends popular items will increase those items’ popularity. Unpopular items will be left in the dust. Such systems can make big hits even bigger, and can lead to an overall decrease in cultural diversity.
Recommender systems can increase the experience of diversity. By drawing attention to items individuals have not found by themselves, they can lead to new experiences. But individual diversity is different from overall diversity. Some systems can increase both individual and overall diversity. Other systems increase individual diversity but, at the same time, prompt consumers to be increasingly similar to each other. Their selections then come from an increasingly narrow range of items [pdf link].
Ownership matters. Given the variety of approaches, outcomes, and absence of clear “best” alternatives, and given the ability of recommender systems to shape the experiences of their users, there is ample room for ulterior motives to become embodied in the system. The incentives for the recommender and the recommendee may be different. The incentives for Netflix in a regime where they deliver physical DVDs (of which they have limited stock) may be to promote the back catalogue. When they deliver movies digitally (as they are about to) there may be no such constraint and they may be more tempted to promote existing blockbusters. The most valuable recommender systems may be those that are independent of producers and vendors.
Transparency matters. The unmarked presence of sponsored items in a recommendation list would be widely viewed as a corrupt set of recommendations, but just as some bookstores charge for premium display sites within the store, so sites on recommendation lists may be sold. Recommendees have a right to know if payola is part of the system.
Recommender systems will displace the filtering role of both reviewers and of publishers. But while bad reviewers and publishers would not be missed, good reviewers and publishers are not only filters; they are also an active part of cultural creation. The impact of recommender systems on these members of creative communities is important.
The word “community” is widely used in conjunction with recommender systems, but they do little to build communities. Their use is essentially an individual, isolated act. Groups and networks are as important in the creation and experience of culture as individuals. Recommender systems will play a role in how culture is experienced, but they are not necessarily a strong force pushing us either towards or away from a healthy culture.
Recommender systems only filter culture, in various ways; the point is to create environments in which artists can prosper.
Nice post. But you don’t talk about the communicative aspect of recommendations, which ultimately requires transparency (i.e., that the recommendation come with a human-readable explanation).
My comments on the subject:
http://thenoisychannel.com/2008/11/21/the-napoleon-dynamite-problem/
I regularly cite Netflix as support for the weak version of the Long Tail thesis, contra to your post last week: when I was living in the States, I was able to find and watch dozens of films I would never otherwise have tracked down — or would have had to expend considerable effort to track down — thanks to Netflix. For me, it transformed my film-watching patterns completely: whereas before I had been constrained by video store selection and avilability, and by recommendations from friends or reviewers, Netflix made it very easy to hone in on films appreciated by others who shared my very specific set of preferences. I would submit that recommendation systems, especially fairly transparent ones (as you have pointed out), are very useful tools for people already adept at negotiating prior recommendation environments (social contacts, newspaper reviwewers), and that enthusiasts/super-consumers of the media in issue can leverage systems like Netflix to effectively filter, and therefore make useful, large back catalogs.
Your point about the “community” claims being off base is exactly right: effective recommendation systems rely on having a large population of users, but if they work properly they actually eliminate the need for the social interactions required to create a “community”, because one can benefit from others’ recommendations without ever interacting with them or knowing that their input forms the basis for the recommendation.
Of course, my enthusiasm for Netflix makes my recent move to Toronto — and the depressing switch from Netflix to Zip.ca — all the more dispiriting. My God, does Zip suck. I’ve finally come to the conclusion that their business model is actually premised on free-riding off of Netflix’s goodwill and providing no meaningful service at all to their customers.
Zip.ca is pure genius!!!! The recommendations engine is completely transparent (item IX). It recommends the most “Available” (i.e. least desirable) titles in its library that I’ve never seen. Actually, I have no proof of the “that I’ve never seen” part. I am 100% confident that no payola is involved in the Zip.ca recommendations. In fact, the recommended movies are such hidden gems that most don’t even have cover images. Without Zip.ca I never would have found the Pilates long tail, never mind the whole Children and Family Television category.
Well I didn’t intend this to be an anti-recommender system for zip.ca, but if they deserve it, well then they deserve it.
Daniel & Picador – thanks for the link and the comments. The points about expert users and about why a system recommends a particular item are dimensions of recommender systems I had not thought about.
Rad – we need to go for a drink sometime.
I wouldn’t classify myself as an expert user, necessarily; an actual expert (a professional film historian, for instance) would presumably have little use for any recommendation system. It’s the next tier down I had in mind — “enthusiast” was the word I used, which I think distinguishes from the casual user and the expert.