4 ms·
Repost of the comment I left on the Github issue: This is going to be a bit of a drive-by comment because I don't have a great deal of time. Apologies in advan
by noelwelsh 13y ago
Repost of the comment I left on the Github issue:
This is going to be a bit of a drive-by comment because I don't have a great deal of time. Apologies in advance.
Firstly, I think you need to define the problem more precisely. I would define it thus: "the algorithm should return a list of sites ordered according to the likelihood of the user finding them engaging." Very quickly you need to decide if you are personalizing this list per-user or not. I'm assuming you're not.
With this definition we can go places because:
1. There is a huge literature on measuring engagement; and
2. There is a huge literature on generating rankings.
For the former we can start by looking here: http://www.dcs.gla.ac.uk/~mounia/ http://www.dcs.gla.ac.uk/~mounia/ (There's lots more work in the field. I just happen to have some familiarity with her work.) For the latter searching for "submodular diversity ranking" will return a lot of papers with complicated algorithms. Yay!
The number one problem I see is that you don't have sufficient information to get a useful measurement of engagement. If you don't have useful feedback you can't learn if what you're suggesting is any good. At the minimum you want to measure clickthrough. Better would be time-on-page (possibly normalised by page length). Solve this problem, even poorly, and we can them look at using this to generate rankings.
As for generating rankings, while the submodular optimisation approach has some very nice properties (e.g. it maintains diversity in the suggestions) there are simpler methods to start with. Let's say we have some measure of engagement. We can put a confidence interval on this measure using standard maths. Then just display the sites that have the higher upper bound on the confidence interval. You should decay the measure over time, widening the confidence interval, to allow for changing interests.
Hope that's useful.