Sep 24 2007
links for 2007-09-24
-
How come I’ve never seen this before? He gives three rules, but only one is important IMHO: “Only solutions that produce partial results when partially implemented can succeed.” I think that’s true for other systems too c.f. agile development
-
Details of the 7.04% solution to the Netflix Prize. Regularised SVD with biases, postprocessed with kernel ridge regression
-
RBMs slightly outperform carefully tuned SVD models
-
Long post with lots of gruesome but readable details of SVD approach to the Netflix data
-
A Bayesian approach to counter overfitting of SVD caused by data sparsity (as in the Netflix data)
-
Uses Principal Component Analysis extended with variational Bayesian learning
-
Expression of Bayesian concepts in an extension to OWL
-
A vicabulary for Bayesian concepts
-
More on representation of Bayesian networks in RDF with a comparison between it and BayesOWL
-
Kendall Clark: “I anticipate that, at some point, we will talk about, say, an RORB (RDF, OWL, Rules, Bayes) platform for Semantic Web development.” IMHO OWL isn’t part of the petatriple future of the semweb. Nor is SPARQL…
-
Interesting points, although I think he missed the biggest reason: don’t rewrite. It always costs more than you think, especially when you have years of obscure bugfixes and workarounds that you have to reimplement.
