Factorization Machine: Held-Out Predictions with Latent Interactions (2D)
Held-out cells plotted as predicted against true value for the additive model and the factorization machine, with test error by latent dimension. Same shared model and checks as the 3D page.
Predict ratings of cells nobody has observed. An additive model knows only that a user is generous and an item is popular; a factorization machine adds a low-rank pairwise interaction between latent vectors. Change the true interaction rank and strength, the noise, the share of observed cells, the latent dimension and the regularisation, and compare held-out error. A Blender-built plotting easel with the O(kn) identity, gradient and closed-form baseline checks built in.
Held-out cells plotted as predicted against true value for the additive model and the factorization machine, with test error by latent dimension. Same shared model and checks as the 3D page.
2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install