Log in
| 1 | ||
| 1 | ||
| 1 | ||
| 1 |
In the pre-tailwind version of the front end there was an upvote button. All it did was add a video to a liked playlist. Now there is both upvote and downvote and also upvote and downvote on thumbnails. It recommends and or disrecomends the video against your history in both directions.
In a way a thumbnail upvote functions a bit like youtube's watch later, and it will behave more like that once I modify the way history works. So if you see more than one video you want to watch upvote the thumbnails and then watch one of them.
The next step I want to take is change the way I handle history so it has a scalar value associated with it. Then a downvote can live in your history and continue to disassociate with history going forward and an upvote will be able to more significantly impact recommendations.
I think doing this plus the recent clamping of magnitudes in both rows and columns of the suggestion matrix, plus log clamping which is some math I won't go into, plus the addition of a few random videos to all suggestions; I think this is going to make gvid suggestions really good.
The column clamping means that a video can only become so recommendable and beyond a certain point will get recommended harder within its own ecosystem instead of everywhere.
The row clamping means, and this is the one I did more aggressively, that no one video can over drive your recommendations.
The log clamping without showing too much math basically besides setting hard limits on some of the row and column magnitudes also tones down their distribution.
But really the upvotes and downvotes should do a lot instead of just rewarding click through. No shame in downvoting something. These aren't reddit downvotes. All a downvote does is kick something back into its own ecosystem.
