224
T. Hanser et al.
Fig. 11.7 Different aspects to take into account when considering the quality and quantity of the
information available to the model to make a prediction. The left configurations of neighbours are
assumed to induce a more reliable prediction
Fig. 11.8 Distance to model (a) and information density (b) can be used to evaluate the reliability
of a prediction based on the amount of supporting evidence for a given query compound
kernel function based on the similarity between the query and the training data
points [27]. The kernel density estimations allow to approximate a data density
map in the descriptor space (Fig. 11.8b).
The role of the reliability is to inform the end user about the strength of the
supporting information; the more supporting evidence available, the more reliable
the prediction, and the more confidence we can have in this prediction. Reliability
is usually expressed as a quantitative value and calibrated between 0 (there was
no relevant information to support the prediction) and 1 (the predicted compound
was known to the model). Values between 0 and 1 give an indication of the level of