Fig 1.
Creation of three-way interaction effect columns.
Fig 2.
Given a full matrix (a), we reduce it to the indices of non-zero entries (b), then the compressed difference between these (c). Arrows represent transitions between different representations.
Fig 3.
Running time on an increasing number of threads.
Note that performance initially decreases with only a small number of threads on a second NUMA node. Tests were performed using two Intel Xeon Gold 6244 CPUs.
Fig 4.
Overall receiver operating characteristic curves for all p = 4, 000 data sets.
Table 1.
Overall AUROC for simulated data sets.
Fig 5.
Running times for each method using large p = 4, 000 data sets.
Note that time is shown on a log scale.
Table 2.
Mean time taken (s) for simulated data sets.
Fig 6.
Receiver operating characteristic curve comparing fraction of reported effects that are true positives as the predicted strength varies in the p = 4, 000 simulations.
Fig 7.
Comparison of methods when 3-way interactions are present.
(a) ROC curve for full fit of all 3-way interaction data sets. (b) Time taken for each method. Note that Pint and Pint (hierarchy) are including three-way interactions, whereas glinternet and WHInter include only two-way.
Table 3.
AUROC for simulated three-way interaction data sets, separated by the type of effect.
Table 4.
InfectX most significant proposed effects.