A value-consistent dimensionality reduction model for electronic tongue data

High-dimensional data are always a great challenge for post-processing. This paper proposed a new computational model for dimensionality reduction toward the huge data sampled from the multi-sensor array in voltammetric electronic tongue. It employs a linear superposition model of noise for a set of basis functions to describe the serial current responses of cross-sensitive multi-sensor array. And the coefficients of the basis function set, with a much lower dimensionality, are used as the reduced feature data for a sample. More importantly, the values of those coefficients show high consistency with the responses of electrodes, namely, larger currents measured may result in greater coefficient values, and vice versa. ing to the theoretical analysis and practical results of the experiments, besides the value-consistency which is never supported by existing approaches, the proposed model also performs very well in the three key factors for evaluating an approach to dimensionality reduction (i.e. reduction rate, reconstruction error, and complexity): (1) strong ability in dimensionality reduction and feature preservation—the reduction rate reaches as high as circa 280–420, and the average reconstruction error is circa 1–5%; (2) linear time complexity—it is applicable to most huge datasets.