Implementing PLS for distance-based regression: computational issues (Q964602)

An algorithm is described for distance-based partial least squares regression. The proposed technique allows to calculate the principal coordinates avoiding explicit Euclidian configuration calculations which, for large datasets, are computationally expensive. Details of computer implementation are discussed for moderately large samples (of size \(\approx 10^4\)) when the distances matrix can be several gigabyte-sized. An application is considered to the problem of rate-making in automobile insurance.