Abstract
Many state-of-the-art recommendation systems leverage explicit item reviews posted by users by considering their usefulness in representing the users' preferences and describing the items' attributes. These posted reviews may have various associated properties, such as their length, their age since they were posted, or their item rating. However, it remains unclear how these different review properties contribute to the usefulness of their corresponding reviews in addressing the recommendation task. In particular, users show distinct preferences when considering different aspects of the reviews (i.e. properties) for making decisions about the items. Hence, it is important to model the relationship between the reviews' properties and the usefulness of reviews while learning the users' preferences and the items' attributes. Therefore, we propose to model the reviews with their associated available properties. We introduce a novel review properties-based recommendation model (RPRM) that learns which review properties are more important than others in capturing the usefulness of reviews, thereby enhancing the recommendation results. Furthermore, inspired by the users' information adoption framework, we integrate two loss functions and a negative sampling strategy into our proposed RPRM model, to ensure that the properties of reviews are correlated with the users' preferences. We examine the effectiveness of RPRM using the well-known Yelp and Amazon datasets. Our results show that RPRM significantly outperforms a classical and five state-of-the-art baselines. Moreover, we experimentally show the advantages of using our proposed loss functions and negative sampling strategy, which further enhance the recommendation performances of RPRM.
Introduction
Modern recommendation systems increasingly rely on textual reviews to bridge the gap between user preferences and item characteristics. However, two significant challenges persist. First, the Noise Problem suggests that not all reviews are equally valuable; a brief, outdated review often carries less weight than a recent, detailed one. Second, the Information Adoption Gap highlights that users process information differently—some prioritize detailed analysis (central route), while others rely on quick cues like ratings or recency (peripheral route).
To address these, we propose the Review Properties-based Recommendation Model (RPRM). RPRM treats review "meta-data"—such as age, length, and sentiment—as primary signals to determine a review's usefulness, personalizing the recommendation based on how individual users prioritize these properties.
The RPRM Architecture
The RPRM architecture is designed to capture the interplay between review content and its associated properties through a dual-tower approach for user and item modeling.
Key Components:
- Review Property Encoding: Review text is embedded using semantic models (like BERT) and then augmented with six specific properties: Age, Length, Rating, Sentiment Polarity, Helpful Vote Count, and Predicted Helpfulness.
- CNN-based Processing: Local features are extracted from these property-encoded embeddings using convolutional layers.
- Review Property Attention: This mechanism learns weights for each property, identifying which ones are most predictive of a user’s interaction with an item. This allows the model to generalize to new reviews by focusing on universal signals like "recency" rather than just specific keywords.
- Novel Learning Schemes: To align the model with human decision-making, we introduced PropLoss (loss functions that reward user-item agreement on property importance) and PropSample (a negative sampling strategy that selects items with similar review properties to those a user typically prefers but did not choose).
Experimental Evaluation
We evaluated RPRM on two large-scale datasets: the Yelp Challenge (Round 13) and the Amazon Product Dataset (across six categories). Both datasets were pre-processed to ensure a minimum of five reviews per user and item to mitigate extreme cold-start effects.
Performance Results
RPRM was compared against six baselines, including Matrix Factorization (BPR-MF) and neural models like DeepCoNN and NARRE. As shown in the table below, RPRM consistently achieved superior results across all metrics (P@1, P@10, R@10, and MAP).
On the Amazon dataset, RPRM achieved a MAP of 0.0378, a significant improvement over the best-performing baseline, NARRE (0.0279). Ablation studies further revealed that "Age" and "Predicted Helpfulness" were the most influential individual properties across both datasets.
Conclusion
By explicitly modeling review properties, RPRM provides a more nuanced and generalizable approach to recommendation. Our findings demonstrate that matching a user's information processing style to an item's review profile significantly boosts accuracy. Future work will explore expanding the property set to include geographical data and leveraging learned property weights to personalize the user interface by highlighting reviews that match a user's specific preferences.
Summary updated