Fair point, I mentioned it because I was thinking they combined the LSA with some form of classifier that got trained as you marked messages as spam (or told it when it misclassified), I'm not sure what Apple's full implementation was though.
Yup - it seems like they used Adaptive Latent Semantic Analysis as a dimensionality reduction technique to represent knowledge[ 0]. LSA helped them use the filtering technique on the device itself as it was not heavy.
