Adversarial Disentanglement Enables Robust Purchase Intention Prediction
IEEE Access · 2026 · 14 · 87192-87212
DOI: 10.1109/access.2026.3699389 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Wei Wang und 2 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
This paper addresses purchase intention prediction in imbalanced tabular e-commerce data. A primary technical constraint is representation entanglement, where observed session features mix latent intent with routine behavioral patterns. We propose the Disentangled Expert Mixture-of-Experts Adversarial Network (DEMA) to separate these components. The architecture utilizes a Gradient Reversal Layer (GRL) as a regularizer to decompose input features into an intent-specific vector ( $Z_{i}$ ) and a purified, intent-agnostic behavioral vector ( $Z_{c}$ ). The gating network uses $Z_{c}$ to route samples to specialized experts, while the experts utilize the concatenated $(Z_{c}, Z_{i})$ for final classification. Performance was evaluated on three datasets. For the Online Shoppers dataset, DEMA achieved a PR-AUC of 0.915 and an F1-score of 0.930. For the Olist satisfaction dataset, it yielded a PR-AUC of 0.915 and an F1-score of 0.915. On a large-scale YOOCHOOSE subset (1,000,000 samples), the model maintained a PR-AUC of 0.824 and an F1-score of 0.795, outperforming GBDT and Transformer-based baselines in skewed class distributions.
Zitieren
APA
Wang, W., Xu, L., & Qu, M. (2026). Adversarial Disentanglement Enables Robust Purchase Intention Prediction. IEEE Access, 14, 87192-87212. https://doi.org/10.1109/access.2026.3699389
Vancouver
Wang W, Xu L, Qu M. Adversarial Disentanglement Enables Robust Purchase Intention Prediction. IEEE Access. 2026;14:87192-87212. doi:10.1109/access.2026.3699389
BibTeX
@article{Wang2026,
author = {Wang, Wei and Xu, Liqing and Qu, Min},
title = {Adversarial Disentanglement Enables Robust Purchase Intention Prediction},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {87192--87212},
doi = {10.1109/access.2026.3699389}
}
EUVIMED