100 000 sources from our spectroscopic test dataset, the F₁ score is calculated from 10 000 nearest neighbours in 1-D (top) and 10-D (bottom) feature spaces, plotted against the mean 1-D feature (left) and mean classification probability (right) for those neighbours. Sources are binned per pixel and colours combined proportional to how many of each class are in that pixel bin. The brightness corresponds to the total source count in that pixel on a logarithmic scale. A small sample of individual data points with their error bars are also overlaid.