Taking into account the recency in the suggestions of videos to compare

[glerzing]

The suggestions API is currently optimized to reduce the global uncertainty in the comparisons. But they do not consider the fact that an entity was recently compared or not.

It has a higher cognitive cost for contributors to compare videos that they have not watched for a long time. They have to recall the content of the video (or the general feeling watching it), or to watch it again. The comparisons should also be less accurate, because the contributors may not have the videos freshly in mind.

Some research exist about the rate at which the memory of something decreases (https://en.wikipedia.org/wiki/Forgetting_curve). But to avoid a curve that decreases too steeply, I suggest a model of the form: score <- score * 1 / (1 + t)^α, with 0 < α < 1 chosen arbitrarily and t the number of days since the last comparison involving this entity.

If trying to reduce the cognitive cost of making comparisons, we could also advantage videos that have a similar length.

[glerzing]

I tried the videos suggestions API. And I think the processing speed is not the main problem. It can suggest very badly-rated videos. Surprisingly, all the top 10 videos that I got from the API as the user1 had a score below -10, I don't know why. Minimizing uncertainty can't be the only criteria for choosing videos to compare. Besides, the code is complicated to understand an verify.

On the other side, the current frontend function getVideoForComparison that mostly picks random videos from the rate-later list works pretty well (at least if people clean their rate-later list regularly, unlike me).

So, I'm pretty reluctant to change it, unless we make more substantial modifications to the suggestions API.