Description
The growth of modern high-throughput instrumentation, which makes avaialble the relatively rapid characterization of samples by often multiplatform and/or multidimensional approaches, results in data that, if not always "big", anyway have a high extent of granularity, with different degrees of similarity and a general inhomogeneity due to many sources of variation, other than those of interest. As a consequence, linear approaches often result inadequate both for regression and classification problems. A possible and obvious solution is to resort to non-linear modelling strategies, possibly where the degree of non-linearity can be tuned depending on the specific characteristics of the data set and kernel-based models, such as SVM or kernel-PLS(-DA) (Rosipal, 2003) are often used for such purpose. On the other hand, local approaches (Centner and Massart, 1998), which build opportunistic models based only of a subset of instances being most similar to the sample(s) to be predicted, represent a valid tool to achieve high accuracy and to take into account the stratification of the data set. However, traditional approaches to local modeling, although having been already used successfully in various domains, have some drawbacks, such a not always clear validation strategy, the lack of specific interpretation tools, their suboptimal performances in the presence of highly irregular distributions and their being almost always limited to single data matrices.
In the present communication, different strategies for dealing with non-linearities or different degrees of heterogeneities for modelling spectroscopic data will be discussed. In particular, new approaches to local modeling will be presented, which try to address some of the limitations mentioned above, and that can be generalized to the cases when the collected data have a multi-way or multi-block structures.
