Monday webinar - The black hole effect in MCR with Raffaele Vitale

In the domain of bilinear Multivariate Curve Resolution (MCR), for instance, the use of Alternating Least Squares (ALS) has been ubiquitous since the very first development of one of the workhorse algorithms proposed for such a purpose, MCR-ALS, in the 1980s. Less attention has been paid to the limitations MCR-ALS can suffer from in case the handled data structures exhibit certain characteristics violating specific distributional assumptions that should be ideally fulfilled when ALS-based approaches are concerned. As an example, it has recently been highlighted how, in the presence of minor components underlying the datasets at hand and if the number of analysed data points is relatively large, the weight or relevance of those that may be essential for a MCR-ALS resolution might become too low for guaranteeing its correctness.

This seminar aims at offering a comprehensive perspective on this particular deleterious phenomenon which has been lately denominated the black hole effect. A distinctive focus will be put on the analogy it has with the biased calibration of univariate and multivariate regression models that may occur when high-leverage outlying observations are collected and processed. Most importantly, two solutions to successfully overcome this effect, namely data pruning and object weighting based on a measure of essentiality for the sake of curve resolution, will be discussed.