DD-ComDim: A data-driven multiblock approach for one-class classifiers

Abstract

An attractive alternative to solve authentication challenges is the combination of fingerprint analytical techniques with multivariate methods of one-class classifiers (OCC). In this proof of concept, we propose a novel multiblock method for OCC that emerged from the association of the common dimension (ComDim) analysis with the dual data-driven, which allows to calculate errors of misclassification based on the orthogonal and score distances with a subsequent determination of their cut-off levels. The applicability of data-driven – common dimension (DD-ComDim) analysis was verified for the authentication of diesel S10 (10 ppm of sulfur) against S500 (500 ppm of sulfur) using two low-field 1H NMR datasets: medium-resolution (MR-NMR) and time-domain NMR relaxometry (TD-NMR). The performance of the DD-ComDim (mid-level data fusion) was compared with data-driven – soft independent modeling of class analogy (DD-SIMCA) for each data set separately and concatenated (low-level data fusion). The results suggest that this novel method can improve the quality and efficiency of the model in authentication when compared to a traditional method. In addition, it demonstrated that the threshold of separation between target and non-target classes was more evident with the mid-level data fusion approach. Other applications are currently underway, intending to verify the method's applicability with other matrices and analytical techniques to confirm the reliability of the DD-ComDim.