Quantifyingthe Bias-Variance decomposition in property predictions on Materaisl Science
Most machine learning (ML) applications in quantum-chemistry datasets rely heavily
on a single statistical error parameter such as the mean square error (MSE) to evaluate
their success or failure. However, this approach has limitations or can even yield incorrect
interpretations. Here, we report a systematic investigation of the two components of
the MSE, i.e., the bias and variance, using the quantum-chemistry QM9 dataset as an
example. To do that, we experiment with three state-of-the-art descriptors, namely (i)
Symmetry Functions (SF, with two-body and three-body functions), (ii) Many-body Tensor
Representation (MBTR, with two- and three-body terms), and (iii) Smooth Overlap of
Atomic Positions (SOAP), to evaluate the prediction process’s performance using different
numbers of QM9 molecules in training samples and the effect of bias and variance on
the final MSE. Overall, low sample sizes are related to higher MSE. Moreover, the bias
component strongly influences the larger MSEs. Furthermore, there is little agreement
among molecules with higher errors (outliers) across different descriptors. According to
the obtained results with the distribution of MSE (and its components bias and variance)
and the appearance of outliers, it is suggested to use ensembles of models with a low bias
(in the case of QM9, the best combination uses two versions of MBTR) to minimize the
MSE, more specifically when using a small number of molecules in the training set.