mpowell wrote:To answer your question: currently Criticker regenerates a set of TCIs every time you rank a new movie, or go on another user's page. Also, every two weeks we do a large set of regenerations for all users.
Plans are underway to regularly schedule full TCI & PSI recalculations (or at least, full recalcs for frequently-active users). The first step was to get faster hardware, which we've done, and the next step is to see how much we can push it, which we've got planned.
Thanks for this info. This sounds pretty good.
mpowell wrote:Thanks for your confirmation that a fresh set of TCIs & PSIs does improve results! That's good to know.
Indeed. The odd thing is that you don't really *
think* that the recommendations are better because your TCI's don't *s
eem* to be very different. My top 10 TCI's were still in the 1.95 - 2.20 range with the % common movies at 36. I think the improvement gain came from all the TCI's *
beyond* the top 10.
Another thing I've noticed that surprised me after I regenerated all the TCI's, was that the accuracy of my PSI's went up when I lowered the % of common movies.
At 50% common movies I got: 32% PSI, 23% IMDB
At 30% common movies I got: 34% PSI, 19% IMDB
At 10% common movies I got: 39% PSI, 18% IMDB
Lower % gives higher PSI compared to IMDB.
This surprised me a bit because my top 10 TCI's at 50% have a lot more movies to go on and their taste *
seems* to be better because you have a lot more common movies than the TCI's at 10%. It *
seemed* as if the 10% TCI's weren't as valuable. But apparently that is not the case. I think the reason for this is that at 50%, Criticker has a lot less data available to use compared to when you use 10%. So probably the PSI's get less accurate, especially for the less common movies because there aren't enough users available to use data from.
Because of this result, I then tested 20% and got something weird: 33% PSI and 23% IMDB. Maybe I made a mistake there, or maybe my numbers aren't as reliable as I thought
. I'll have to test that again some time.So........... what I'm currently thinking, is that your top 10 TCI's are more valuable to the user at 50% (me) when you're casually looking for a good movie,
whereas for the core business of PSI generating you really need to lower the % to get the best results.
Now... this makes me wonder... how good are the PSI results for people with 1500 rated movies (unlike my 400)? For them 10% is 150 common movies and 50% is 750 movies, right?. How many users are there on Criticker that have 750 movies common with someone else? Don't the recommendations get worse when you've rated so many movies because you can't lower the % of common movies lower than 10%? If a user with 1500 rated movies sets his % at 50, doesn't he get a 'N/A' TCI for all but a few users?
Maybe you shouldn't demand a % of common movies, but an absolute number of common movies? For me, 10% gave the best results = 40 movies, whereas 50% gave the worst results (200 movies).
Something to think about... If you have some further insights into this, I'd gladly hear it

.