Vector-based categorization analysis with improved principal component generated reference

TL;DR

This paper introduces an improved vector-based classification method using principal components to better identify clusters in large, variable datasets, demonstrated on molecular junction measurements.

Contribution

It proposes using the first principal component of the correlation matrix as a reference trace for enhanced clustering in vector analysis.

Findings

Principal component-based reference trace reveals two distinct groups in data.

Method improves cluster separation compared to traditional blank tunneling reference.

Supports objective hypothesis formulation in large data set analysis.

Abstract

With automated measurements, large data sets can be assembled with relative ease. Often these data sets have a large degree of variance. Nonetheless, we hope to find groups within the data set with shared distinguishing characteristics. Recently Albrecht, et al demonstrated a multi-parameter vector-based classification analysis which provided a powerful means to find clusters within a data set of conductance versus displacement traces measured from single molecular break junctions (Lemmer, et al, Nature Communications, 2016, 7, 12922). An idealized blank tunneling trace was used as a reference to calculate three vector variables. The authors suggested a more appropriate reference can be chosen to better form clusters in the data. Here we propose using a principal component of the correlation matrix calculated from the data set to construct a fast and strategic reference trace. Principal components form an orthogonal basis set in units of the independent variables comprising the correlation matrix. The first principal component, the one corresponding to the largest eigenvalue of the correlation matrix, points in the direction of largest variance in the data. We used both a blank tunneling reference trace and a reference trace constructed from the first principal component of the correlation matrix to analyze a set of 4,4'-bipyridine single molecular break junctions. The blank tunneling reference trace produced a single elongated cluster in the vector analysis plot whereas the reference trace constructed from the principal component distinguished two arms in the cluster. We hypothesize that these two arms are due to groups of traces with either long or short molecular plateaus. Statistically based sorting algorithms, like vector-based categorization techniques, provide an objective framework to formulate such hypothesis and aid in designing new and pointed studies to verify them.