Abstract
Traditional recommendation models often treat all user reviews with equal weight, failing to account for the varying "helpfulness" of different reviews to different users. This paper introduces the Review Properties-based Recommendation Model (RPRM), which integrates review metadata—such as age, length, and sentiment—into an attention-based neural architecture. By explicitly modeling these properties and employing novel loss functions like PropLoss, RPRM significantly outperforms state-of-the-art baselines on large-scale datasets from Amazon and Yelp.
Introduction
While modern recommender systems increasingly leverage review text to understand user preferences, they frequently overlook the context in which a review is consumed. In practice, not all reviews are equally influential; a user may prioritize a recent, detailed review over an older, brief one. Existing attention-based models attempt to learn review importance from text alone, but these often fail to generalize or capture the diverse "information adoption" styles of users. Some users follow a "central route," focusing on logical content, while others follow a "peripheral route," relying on quick cues like star ratings or helpfulness votes. The authors propose bridging this gap by explicitly modeling review properties to determine their impact on recommendation accuracy.
Methodology
The RPRM architecture is a dual-tower neural network designed to model users and items through their associated reviews. As shown in Figure 1, the model consists of four primary components:
- Review Property Encoding Layer: Reviews are embedded using BERT (768 dimensions) and weighted by six normalized property scores: Age, Length, Rating, Sentiment Polarity, Helpful Vote Count, and Predicted Helpfulness Probability.
- Review Processing Layer: Convolutional Neural Networks (CNNs) extract local features from the property-encoded review vectors.
- Review Property Attention Layer: A dot-product attention mechanism learns which properties (e.g., recency vs. detail) are most salient for a specific user or item profile.
- Prediction Layer: The final representations are combined with standard user/item ID embeddings to predict preference scores.
Beyond the architecture, the authors introduce PropLoss, a set of loss functions (PropLoss_uu and PropLoss_ui) that reward the model when a user and an item "agree" on property importance. Additionally, PropSample is a negative sampling strategy that selects items with similar review properties to those a user has liked, forcing the model to learn more nuanced distinctions.
Experimental Results
The model was evaluated on the Yelp (restaurant category) and Amazon Product datasets against six baselines, including BPR-MF, DeepCoNN, and NARRE.
As detailed in Figure 2, RPRM consistently achieved superior accuracy. On the Amazon dataset, RPRM reached a Precision@1 of 0.0223, significantly outperforming the strongest neural baseline, NARRE (0.0175), and traditional BPR-MF (0.0053). Key insights from the experiments include:
- Property Significance: Adding review properties to a baseline BERT model ("No-Prop") significantly boosted performance, confirming that metadata provides vital signals that text alone cannot capture.
- Dominant Features: "Age" and "Predicted Helpfulness" were the most effective properties for identifying useful reviews across both datasets.
- Personalization: The attention mechanism successfully identified varying user profiles; for instance, some users rely on review length while others prioritize recency.
Conclusion
RPRM demonstrates that explicitly modeling review properties and user-item agreement on those properties leads to more accurate recommendations. The integration of BERT with property-aware attention provides a robust framework for capturing complex information adoption behaviors. Future work may explore expanding the property set to include geographical data or linguistic style markers, and applying these importance scores to personalize the presentation of reviews in user interfaces.
Summary updated