Determine the role of structural recovery in ensemble gains

Determine whether the recovery of higher-order structures by Cluster-TNN explains the predictive gains obtained through ensemble inference.

Background

Cluster-TNN trains Topological Neural Networks by partitioning an input graph, sampling groups of clusters, reconstructing the induced subgraph, and applying the selected higher-order lifting only within each mini-batch. The paper establishes theoretical conditions under which higher-order structures can be recovered across repeated cluster recombinations and reports that ensemble mini-batch inference produces higher mean test metrics than single-pass inference in all 21 controlled comparisons.

However, the paper does not determine whether this structural recovery is the mechanism responsible for the observed ensemble-inference gains. Establishing that causal or explanatory relationship would clarify whether improved predictions arise from exposure to additional higher-order contexts, from ensembling across sampled views more generally, or from other effects of the batch-before-lift procedure.