Determine whether alternative aggregation rules preserve ordinal structure

Determine whether an aggregation rule other than the fixed deterministic per-attribute-type choice used in the transaction-pattern attribute database preserves ordinal structure better, particularly for education attributes, by evaluating alternatives such as threshold-based, per-attribute-type, or prior-weighted aggregation.

Background

The pipeline aggregates attributes inferred for transaction patterns into user-level profiles using a deterministic rule selected according to attribute type. The paper reports that this rule maintains education ordinal mean absolute error but loses the underlying ordering signal, as measured by Spearman correlation.

The authors explicitly leave unresolved whether another aggregation strategy—such as thresholding, a different per-attribute-type choice, or prior-weighted aggregation—could preserve ordinal information more effectively. Resolving this issue would clarify how much of the observed performance depends on the aggregation mechanism rather than on the pattern-level inference itself.

References

The pattern-to-user aggregation rule is fixed at a deterministic per-attribute-type choice without ablation; on education this fixed choice keeps ordinal MAE stable while losing the ordering signal (Section~\ref{sec:results}), and whether a different rule preserves ordinal structure better is left to future work.

— From "Who Is This User?" to "What Does This Purchase Mean?": A Deployed Pipeline for Semantic User Profiling at Bank Scale  (2609.19928 - Mitsuhashi et al., 17 Sep 2026) in Section Discussion and Conclusion, paragraph beginning “The evaluation of Section~\ref{sec:results} carries five limitations”

The breadth-dispersion statistic is a new and systematic failure no earlier diagnostic surfaced: real buyers are substantially more heterogeneous in breadth than the simulator, and the statistic lies outside band in three of four cells while passing cleanly in the fourth. That axis is one the variance-injection weight does not touch, since it diversifies which brands an agent buys rather than how many. We therefore report partial transfer on one untouched axis together with a previously undetected limitation on another, and treat a breadth-generating mechanism as future work.

— Testing, not presuming, adequacy: calibrating generative social simulators against emergent network structure  (2609.24012 - Shao et al., 21 Sep 2026) in Section 5, “A post-hoc statistic-held-out audit weakens circularity and exposes a further limitation”; Discussion, “Scope conditions, and the questions the residual opens”