Geo-Invariant Scoring Lead with Domain-Adversarial Transformers

ICLR 2026 Conference Submission21486 Authors

Keywords: domain adaptation, transformers, lead scoring, adversarial learning, geographic fairness, DANN, sequential modeling

TL;DR: We augment transformer-based B2B lead scoring with domain-adversarial training to achieve geography-invariant representations, reducing regional performance gaps by 12.3% without degrading majority-region accuracy.

Abstract: Predicting B2B lead conversion requires not only modeling long‑range dependencies in richly sequenced customer interactions but also ensuring fair performance across under‑represented geographies. While our DeepScore transformer backbone improved overall AUPR from $0.266$ to $0.360$, it exhibited significant geo‑skew: majority‑region (America) signals dominated feature learning (AUPR $0.474$), leaving East-Asia ($0.262$) under‑served. To address this, we embed a Domain‑Adversarial Neural Network (DANN) module into DeepScore’s architecture. A gradient‑reversal layer connects a geo‑discriminator to the shared transformer encoder, enforcing a minimax game that drives hidden representations to be predictive of conversion yet uninformative of geography. Simultaneously, lightweight geo‑specific classifier heads learn residual region‑nuances without re‑introducing large divergence. DeepScore + geo‑DANN achieves a $4.3 \%$ relative gain in macro‑AUPR and reduces inter‑region AUPR gaps by up to $12.3\%$ , all without degrading America accuracy. To our knowledge, this is the first demonstration of adversarial domain adaptation in large‑scale B2B lead scoring, offering a scalable path to equitable, high‑fidelity predictions across heterogeneous markets.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 21486