An second thought the rank percentile would be heavily influenced by the amount of ratings given per user.
If you rate only 3 games you think are great with high ratings, then one of them is going to recieve an awful percentile.
The same 3 great games rated by someone with 5000 ratings will have much 'better' percentiles, just because all the crap games.
Really nice! As an extension to the article, I'm also making a recommender, but just colab filtering. But yours looks stellar! And the article is great, compliments!
Need some time to let the like score calculation sink in :-)
I'm going to experiment with the (rating * 2) / 100, seems like a great way to account for the nonlinearity.
Btw don't you divide by 10 instead of 100?
Another suggestion was to take transform the ratings of each user to percentiles, as a measure of how favorite the game is to the user, also seems interesting.
Actually I'm making a small application that shows similarly rated games. For fun tried out Brass: Lancastershire and got as most similar games: Brass: Birmingham, Age of Steam, Food Chain Magnate, Dominant Species
Top 3 for Eastfront: Advanced Squad Leader: Starter Kit #2, Unconditional Surrender! World War 2 in Europe and..... A Victory Lost: Crisis in Ukraine 1942-1943
The rank percentile is a great suggestion and after that take the average rank I guess?
Ratings per year is a possibility.
Complexity: Indeed I wanted to remove some of the more complex games because I don't have time for them. Although on reddit there was a nice discussion that the weight is composed of complexity & depth
The same 3 great games rated by someone with 5000 ratings will have much 'better' percentiles, just because all the crap games.
And all the shades in between.