Yes just one. Adding more latent factors seems to make little difference. A single factor seems to explain most of the polarization among users.
I have run it with multiple dimensions and plotted the results. They are fascinating, even if they don't necessarily improve the algorithm. See the 3D plot in my article on Improving Bridge Based Ranking, the section "3D And Higher-Dimensional Matrix Factorization:
https://jonathanwarden.com/improving-bridge-based-ranking/#3...
That is really interesting! I would have guessed that multiple dimensions would be required, e.g. religious/atheist, conservative/liberal, parent/no-kids, who knows... after all there are a variety of community notes topics.
One possible explanation for why these dimensions don't improve the algorithm a lot is that differences in these dimensions don't cause differences in whether users rate a note as helpful or unhelpful -- at least not beyond what can be explained by the primary latent factor. These other dimensions may contribute to other factors of user behavior -- such as which tweets they like and what users they follow. But once we know a user's left-right polarity factor, these other factors don't make a huge difference on whether or not a user ranks a note as helpful (given they rated the note at all). They do make a difference, but since most of the difference is already explained by the polarity factor, they don't add much to the algorithm.
I have run it with multiple dimensions and plotted the results. They are fascinating, even if they don't necessarily improve the algorithm. See the 3D plot in my article on Improving Bridge Based Ranking, the section "3D And Higher-Dimensional Matrix Factorization: https://jonathanwarden.com/improving-bridge-based-ranking/#3...