Counts of early contact days reduced later contact forecast error for observed pairs in two office data sets

We tested whether counts of early contact days improve a later contact forecast beyond total contact time and five other inputs. We used two public SocioPatterns office data sets, from 2013 and 2015. The first five recorded days gave inputs. The last five days gave the outcome: any recorded contact. We fitted models in one release and tested them in the other, in both directions. Among 2,916 early pairs in 2015, adding contact days reduced the Brier score from 0.16648 to 0.16157. Among 521 early pairs in 2013, it reduced the score from 0.19901 to 0.19399. The relative reductions were 2.95% and 2.52%. All planned sensitivity checks kept a mean gain. Target-node resampling gave a descriptive 95% difference range of -0.00904 to -0.00100 for 2015 and -0.01383 to +0.00432 for 2013, with source fits fixed. Different-department probability error fell in both tests, but ranking became slightly worse in the 2015 target. The result concerns two short releases from the same office setting and pairs observed early. It does not establish a causal effect, a better list of department bridges, or performance in a new workplace. The package preserves raw data, source hashes, the local plan, all forecasts, figures, and reproduction code.

temporal networks · human contact · contact days · link prediction · workplace · reproducibility