The main purpose of these experiments was to show the performance of other reconstruction methods different than the one using the decomposition matrix .
Except for Glass that contains not enough data for this type of approach to be interesting, decomposing the problem into subproblems is fruitful and elaborated reconstruction methods are better than simple ones. k-NN and SVM provide the best results, but the latter method required a lot of hand-tuning (e.g. for each database, the best values where obtained for different kernels or different parameters for these kernels). It is also interesting to note that decision trees (DT) or neural networks (MLP) are not significantly better than the best linear reconstruction methods (Ho-Kashyap).