I tried out the Steam Interactive Recommender and cranked the dial to MAXIMUM NICHE.
It suggested to me:
1) Vanquish (already on wishlist, already own on PS3)
2) Deadly Premonition (no thank you, but I understand why)
3) System Shock 1 (makes sense)
4) Xanadu Next (have long considered it)
5) Castlevania Lords of Shadow Mirror of Fate (i've heard it's bad).
Further down the list, it recommended me AAA games like Dark Void, Lost Planet 1, Singularity (ha!), and DMC HD collection. This tells me that Steam sees "niche" as games that sold poorly or get less external traffic. Which makes sense from a computer's point of view, but stuff like Vanquish, Mirror of Fate, Dark Void, Lost Planet, and DMC HD Collection just sold like shit, they aren't niche and didn't receive niche coverage externally.
They likely draw the line as to what niche is at too high a level (and should use
Wok 's algo...or at least be able to slide it further down to reach that).
That said, it also showed me some obscure shmups and a few platformers I hadn't seen before.
I'll say this though - when i cranked the slider the other way (to popular), it gave me a whole bunch of games already on my wishlist, especially ones that rank highly. I think a lot of people would see that as a failure of the system, but I see it as the opposite; the machine learning appears to know almost exactly what I'm directly interested in.
That said, what it needs is another slider; the algo should take the person's wishlist into account but there should be a slider that allows the user to pick the strength of the wishlist's listings on influencing the choices. I say this because people have different uses on their wishlists - some people just throw anything and everything that looks vaguely interesting to them on there, some people use it as a very specific purchase list. For the former category, those users would probably be best lowering the influence of the wishlist influence, the latter might as well count them as good as purchased when it comes to influence.
These are things I would suggest implementing:
1) a checkbox beside the games shown on the left (most played list). Unchecking a game means it won't count towards the algo. Some people play games they hate for a long time, sometimes you don't want to play a game like that ever again, etc.
2) There should be a checkbox that says something like "Don't show me things already on my wishlist". Just to de-clutter the list when 10 of the top 20 are already things you plan to buy.
3) Some way to incorporate your entire library
4) Some way to incorporate your wishlist
5) Some way to mark a game in the suggestion list "I'm definitely not interested in this" and then incorporate that back into the algo for you and other users. MORE DATA.
All in all, a super cool start to a possibly very useful system.
EDIT: i think a big weaknes of the system is that it seems to only take into account play time, rather than "liked time". Theoretically, i wouldn't have spent 40 hours playing GTA4 if I hated it, but, welp, humans are stupid and I did. Just because I played Skyrim for 5 times as much time as Dead Space doesn't mean i like it more. And if it's just matching it up with people who also spent 100 hours on Skyrim and 20 hours on Dead Space, that doesn't mean we like the same stuff, it just means we played the same games. It's the multiplayer vs singleplayer time fallacy - you could probably nearly directly correlate a player's time in a multiplayer game with the amount they were engaged by it, but a lot of people play single player stuff to the end regardless of quality (unless they absolutely hate it), so long games just mean longer play times...but I might have LOVED, say, Gone Home for all its 2 hours.
I'm not sure how one would account for this. I would guess that single player games have really distinct MODE averages, with a cluster at the low hour count (1 - 2 hours) and then another high mode at the MEAN completion time. How does one filter that out to determine like-itude?