Figure 1
From: Model selection to achieve reproducible associations between resting state EEG features and autism
Diagram of model robustness evaluation procedure. Given a dataset \(\:\mathcal{D}\), predictive performance metric \(\:P\in\:\left[-\text{1,1}\right]\), similarity measure \(\:S\in\:\left[-\text{1,1}\right]\), and a model with linear coefficients \(\:\varvec{w}\), data is first partitioned into two disjoint sets, \(\:{\mathcal{D}}_{A}\) and \(\:{\mathcal{D}}_{B}\). The model is trained twice – once on each set – and tested on the set on which it was not trained. The two test performances (\(\:{P}_{A}\) and \(\:{P}_{B}\)) are averaged to get the predictive performance \(\:P=\left({P}_{A}+{P}_{B}\right)/2\) of the model. The similarity of the coefficient representations is then quantified using the chosen similarity metric, \(\:S=\text{s}\text{i}\text{m}\left({\varvec{w}}_{A},{\varvec{w}}_{B}\right)\). Model robustness \(\:R\) is calculated as the multiplicative product of predictive performance \(\:P\) and representational similarity \(\:S\): \(\:R=P\cdot\:S\).
