A comparative analysis of Recommender Systems - Amazon and Netflix

Recommendation systems tend to add significant value to a web-store, particularly since a. they can increase traffic towards less demanded items in the store, and b. they can enable you to leverage your existing customer base to "market" to users considering purchases of new goods related to ones other customers with similar tastes have bought in the recent past.

There are however several non-technical aspects one needs to be aware of as one designs recommender or recommendation systems, among them the following:

1. recommendations provided must use a very "light touch" and indeed be relevant, or users will bail (recall that users have very low switching costs in the Internet/web context)
2. it is usually a good idea to tell users on what basis a recommendation is made, to reduce annoyance or increase delight at the use of your systems - e.g. are we recommending a product to you because others like you (similarity defined in any reasonable way - similar profile etc) liked that product? because others who bought the same products you did, bought this other product? because others with the same browsing history as you bought the product in question?... and other similar considerations
3. ease of rating and generating reviews
4. ease of user-base interaction amongst themselves (the more vibrant your user community, the less hand-holding you might need to do, if you have good automated systems to moderate user forums etc)

Also, a key difference here is whether you are selling your own products, or acting as a store-front to others' products - in the latter case, you can of course be much more open and neutral in terms of the user-generated content that is hosted. And quite frankly, I don't think I have seen a good example of the former kind of systems in practice.

That said, on to the topic at hand... Both Amazon Prime Instant Video (APIV) and Netflix Video (NV) have recommender systems built in, but they differ in some important ways:

1. APIV says "people who watched what you just watched" also watched these titles A, B, C... While NV says "people who liked the title you just watched, also liked these titles E, F, G." 
2. APIV as perhaps the largest and most widely used store-front today (sorry EBay) knows a lot about what you've bought, and can leverage that to make recommendations. Of course, APIV doesn't know whether you liked what you bought or watched already in all cases. NV only knows what movies you have seen and which ones you liked (based on what you've rated or said you liked) and can use that as a basis for making recommendations.
3. Both leverage information about user behavior. Neither uses information about the product. For instance, new movies, before anyone has watched them, will have zero stars in both NV and APIV. 

Between the two, you could argue that NV offers the more targeted recommendations primarily because these are tuned to your taste, based on what you've told NV already that you liked. APIV's recs, on the other hand, are much more generic, and are more focused on getting people to watch videos using their free subscription, and arguably less focused on whether this is what people might want to see.

NV has the better store-front for renting/streaming video - the popups on mouse hover over movie icons is something users expect, and it is disappointing to see Amazon doesn't implement yet. But then Amazon has hands-down the better infrastructure at the back-end - don't forget, even Netflix is hosted off Amazon servers. But it is user experience we are talking about here, and NV wins.

It is not hard to envision a world where a recommender system is stripped off a store-front and is run by a third party. This puts various streaming store-fronts on a level playing field where the main focus here is efficient user choice, and potentially, cost. We have less time to waste these days given our busy schedules, and have to choose from a larger variety of products more quickly. A system that helps us make the correct choice more efficiently will drive storefront success. A system decoupled from a store-front owner gives a sense of neutrality and unbiased-ness which is critical to its user acceptance.

Billing becomes more complicated in the latter scenarios however. Currently the user pays the service provider for the service. In the latter case, the user may pay the service provider who kicks back some amount to the middle-man for referring users to their site, or the user has to pay a fee to the middle-man for access to better sites. The former model has similar issues with neutrality and conflict of interest as one saw with Ratings Agencies during the recent Financial Crisis and debacle. The latter... well, it complicates matters significantly. And no one wants to pay more for something they're accustomed to getting for free now. Look at the challenges New York Times and now Washington Post are having at monetizing their digital content.

It is a pity that neither NV nor APIV utilize recommender systems that leverage thematic elements in movies. For instance, one can compute statistically improbable phrases or SIPs within either movie synopses or in movie content (e.g. subtitles) and try to determine other movies with similar content that watchers might like. TF-IDF based systems that do this might not be too difficult to build.

So we didn't mention Hulu in the above. Well, that business model is slightly difficult to comprehend. The most common gripe I hear from even fans and subscribers of the website is that even after paying for premium membership, the user is still subjected to ads. True, the "premium" aspect provides for easier access to full seasons of shows, but people feel cheated that after paying for an "Internet" experience, they are still treated like "ordinary" TV watchers. After all, if you allow time-shifting and place-shifting, you should permit the more discerning users to watch videos the way they want as well, especially if you're charging them.