My understanding is that Kendall Tau is a probabilistic interpretation on the non-parametric test, so the value of +.392 can be somewhat interpreted probabilistically. I think a supervised learning method is likely to obtain even better results. We did something like this to predict the amount of conflict in Wikipedia articles in our CHI2007 paper and was able to get r-values close to .9.

# Aniket Kittur, Bongwon Suh, Bryan Pendleton, Ed H. Chi. He Says, She Says: Conflict and Coordination in Wikipedia. In Proc. of ACM Conference on Human Factors in Computing Systems (CHI2007), pp. 453–462, April 2007. ACM Press. San Jose, CA.

Another experiment: I used the full regression on v1..v5 as a model for skewz ratings. The scatterplot looks a bit better, which makes sense since the R^2 is better. If you think you want to as closely model Skewz ratings as possible, this model could be used for the final scoring — just dot product a site’s SVD v1..v5 values against the coefficients above. This is definitely into semi-supervised territory now (like how BellKor uses NMF and other SVD-like outputs as features in a linear regression against the NetFlix ratings), and sadly not quite as simple and magical as just grabbing the v2 :)

Plot image here and I’m putting it in the post above too.