Methodology for evaluation

[GolanTrev]

We need to differentiate evaluations in a subset of special cases, versus evaluations on realistic datasets.

• 3 experiments related to: missing stops, merging, splitting. Of the type "for users with x, y, z properties, then this problem happens in a, b, c regime of parameters". Just proof of concept, no need for all combinations of everything.

• Think of 1 experiment that is more general. Perhaps we look at a histogram of q (completeness) of an actual dataset, and at some preferential return patterns, and we test a whole population without focusing on special cases.