AdaSeq-MTA: Attribution-Aware Adaptive Neural Computation for Multi-Touch Attribution

Abstract Processing long customer journeys and estimating touchpoint contributions can make neural multi-touch attribution computationally expensive. Touchpoint selection reduces this cost but may change credit allocations even when conversion predictions remain similar. We propose AdaSeq-MTA, an adaptive neural framework that jointly optimizes conversion prediction, credit consistency, and com-putational cost. Guided by a frozen full-journey teacher, the model selects a processing path andretention budget for each journey. It maps selected credits back to the original touchpoints and align stouchpoint and channel allocations, with additional terms for signed contributions and positive-creditmass. A workload-level constraint controls routing, selection, and encoding costs, while confidence-controlled Shapley sampling adjusts attribution effort. On the Criteo benchmark, AdaSeq-MTAretains 66% of touchpoints and reduces inference latency from 18.4 to 13.1 ms per journey, a 28.8%reduction, while the area under the receiver operating characteristic curve (ROC-AUC) changes from0.7758 to 0.7753. Channel-credit drift decreases from 0.029 for prediction-only adaptive routing to0.019. With 100 Shapley samples, touchpoint selection reduces attribution latency from 156.4 to104.2 ms per journey, a 33.4% reduction relative to full-sequence evaluation. These results show that AdaSeq-MTA reduces inference and attribution time with a small change in predictive ROC-AUC, while achieving lower credit drift than prediction-only adaptive routing.