I had fun playing with this, but what I noticed is that the recommendations started well, but as I rated more movies, they became less and less accurate.

Is it possible that whatever technique you are using to model individual users is overfitting to noise in their ratings?