Cross-Domain Review Helpfulness Prediction Using CNN with Auxiliary Domain Discriminators

In June 2018, the NAACL conference featured a paper from Ant Financial’s AI department titled “Cross‑Domain Review Helpfulness Prediction based on Convolutional Neural Networks with Auxiliary Domain Discriminators,” which addresses the challenge of estimating the usefulness of product reviews on e‑commerce platforms.

Because many product categories have limited labeled reviews and suffer from out‑of‑vocabulary (OOV) problems, the authors propose a method that leverages transfer learning to improve helpfulness prediction across domains.

The proposed solution consists of two main components: (1) an improved TextCNN that augments word embeddings with character‑level representations to mitigate OOV issues, and (2) a specific‑shared adversarial transfer‑learning architecture that learns shared representations (h_c) together with domain‑specific representations (h_s for source, h_t for target) while incorporating domain‑discrimination losses (L_s and L_t) to encourage domain‑invariant features.

The model is trained as a regression task to predict the proportion of users who consider a review helpful. Experiments on a public Amazon review dataset covering five domains (Watches, Phone, Outdoor, Home, Electronics) show that the base CNN outperforms handcrafted feature methods, the enhanced CNN surpasses ensemble baselines, and the transfer‑learning framework yields significant gains when target‑domain data are limited.

Further analysis reveals that the transfer‑learning benefit is most pronounced when only 10%–30% of target data are available, while performance converges when full target data are used. The approach has also been deployed in Ant Financial’s anti‑fraud scenarios, and the authors plan to extend it to additional business contexts.