The algorithm was already very good, it is now cutting edge.
Do we use AI ? Not at all !
First thing, we categorized Straight and gay clips on the website.
For Straight, It is impossible to get recommended M/M clips, unless a clip features features a mix (we call them bridge clips)
The second part of the process was harder : we used anonymized buying stats and try to see buying patterns with powerful algorithms.
The result is that the process is now more Mood based than tag based :
- Using tags is like trying to guess what ou like to eat by judging the ingrédients you buy : Salt, flour … You can do a lot of things with that.
- Moods are based on combination of tags : sensuality, nylons/socks … Intense stuff. Instead of flour and salt, we look at Pizza dough directly. It will not tell us if you prefer Pineapple Pizza or a Marguerita, but it lets you discover.
Tags are not gone, we still look at them but they refine instead of guide.
We also started to look at broad categories.
What we did not do :
Sementic recommendation is superior in theory, you look at descriptions / titles, then you give more points to something similar.
The problem we have : the descriptions on the website are not of same quality, some sellers use a few words, others long text. The titles are often series, so you get into a closed system very quickly.
The new system is now live, but we consider it in beta. If you find weird recommendation, tell us
