Commit c8337f

2024-10-22 14:18:02 Doku: Writeup
/dev/null .. projects/anime recommendations.md
@@ 0,0 1,9 @@
+ # Anime recommendations
+
+ This is a project to find similar anime based on the similarity of their reviews. It follows the example of Andrej Karpathy's movie recommendation project.
+
+ I process a dataset of anime reviews and compute their embeddings. Then I average out the embeddings of the same anime. Finally, when I pick the anime, I find the closest embeddings to it, which are considered to be similar anime.
+
+ UMAP visualization of the embeddings shows some interesting patterns, like how Pokemon is very different from everything else, how some franchises also have "islands" of embeddings and how you can roughly see clustering by genres.
+
+ I had to take down the site where you could see all this, but I'll put it here soon.
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9