Past contact data reduced later contact reconstruction loss in two records from one workplace

We tested whether past pair contact data improve daily contact reconstruction when daily co-presence is known. We used two public SocioPatterns records from one French workplace, InVS13 and InVS15. We fitted fixed logistic models on four early target days and tested five later days per record. The baseline used current and past co-presence, a same-department indicator, and current node activity. Adding previous observed-day contact occurrence and duration reduced mean daily log loss by 5.13% and 3.74%. Each test day had lower loss. Mean daily average precision rose from 0.319 to 0.415 and from 0.417 to 0.494. Checks chosen after the planned test added past node contact activity, department-pair indicators, and nonlinear exposure terms. Pair history still reduced loss by 3.90% and 2.34%. A further check used only early-period contact history and these added controls. It reduced later loss by 3.71% and 2.13%. Co-presence covered only 81.5% and 84.9% of test contact pair-days. Model scores apply only to that candidate set. The task uses current-day co-presence and is not a forecast. The two records do not support population inference or a causal claim. The package has code, source provenance, aggregate results, figures, and repeat instructions.

human contact networks · co-presence · contact history · daily reconstruction · temporal networks · SocioPatterns · chronological validation