Graph machine learning for activity scheduling: Location choice set formation

Abstract

This paper represents a first step towards a graph-based, econometric activity modelling framework, incorporating both location and spatial choices into the behavioural process. Current econometric activity based models typically focus predominantly on the temporal aspects of activity scheduling, leaving spatial decisions to agent-based microsimulation frameworks. Whilst recent work has established behavioural frameworks that incorporate location choice into the scheduling process, they have not yet been applied practically as (i) joint schedule optimisation across large numbers of candidate locations is computationally infeasible, and (ii) associated methods to establish individual location-choice sets for this context have not yet been developed. We address this gap by framing location choice set formation as a graph-based node classification problem. Using Graph Neural Networks (GNNs), we predict, for each individual, the probability of visiting different locations across a transport network graph enriched with point-of-interest, land-use, and individual sociodemographic features. We then demonstrate two different sampling strategies to generate choice set from the predicted node probabilities. We validate our approach on a synthetic case-study as well as real-world historic travel data from the Geneva public transport network, showing that GNNs can recover behaviourally plausible location choice sets while remaining computationally tractable.