>>
|
No. 36088
[Edit]
Replies: >>36089 >>36620
>>36087
You wouldn't (/couldn't) actually explore the high-dimensional space itself, the raw embeddings would be used in various fashions. For instance clustering ("find anime having characteristics of both X and Y", "anime like X, Y, Z but not Q"), or use it to do lower-dimensional projections (set dim1 as the trait shared by "X, Y, Z", set dim2 as the trait maximally different between "X, Y, Z" then project onto {dim1, dim2} and explore).
>Genres, demographics
Genre/demographics would implicitly be included as dimensions. Again, this could be done either via traditional (non-ML approach) kind of similar to "recommendation systems" which have well-studied solutions in the form of matrix factorization (e.g. SVD, which would directly give you a set of orthogonal dimensions), or from newer ML techniques (there's probably some fancy way to compute embeddings from ratings, shows that are talked about together frequently, etc.).
>and best of all, recommendations from those with similar taste, are already effective.
But that's exactly the goal – Ideally if there were someone with the same tastes as you who had knowledge of the entire anime corpus, you could use their recommendations. But of course such people often don't exist, and if you do have any friends their tastes may only partially overlap with yours. Which is why if you aggregate together all this info to get a matrix of ratings/recommendations, you can then go from a "local" view to a "global" view to tease out the various features a show could have and how important these features are to each person who rated.
--
I think it should actually be feasible to do this since people have scraped the MAL profiles (UserAnimeList.csv). This gives you a setup basically identical to the netflix prize, I wonder if anyone has ever tried to do stuff with that.
|