Abstract
Modern recommender systems increasingly leverage review text to capture nuanced user preferences and item characteristics. However, standard approaches often process user and item review histories using symmetric architectures, failing to account for the inherent diversity in a user's history compared to the specificity of an item's reviews. This paper introduces the Asymmetrical Hierarchical Network (AHN), a novel framework that utilizes an asymmetric attention mechanism to dynamically select relevant features from a user's heterogeneous review history based on the target item. Evaluated on ten real-world datasets, AHN consistently outperforms state-of-the-art baselines while offering high interpretability.
Introduction
The core challenge in review-based recommendation lies in effectively extracting signals from noisy textual data. A fundamental observation is the structural difference between user and item review sets: a user's history is typically heterogeneous, spanning various product categories and features, whereas an item's reviews are homogeneous, focusing on a specific set of attributes. Traditional symmetric models struggle to isolate the relevant "signal" within a user's diverse history when predicting their interest in a specific new item.
Methodology: Asymmetrical Hierarchical Network (AHN)
The AHN architecture processes textual data through a multi-level hierarchy, progressing from words to sentences, and finally to reviews to learn comprehensive user and item embeddings.
The model's primary innovation is the Asymmetric Attentive Module:
- Item-Side Self-Attention: Since item reviews are focused, a self-attention mechanism with a non-linear gating function identifies core item attributes across all reviews.
- User-Side Co-Attention: This module uses the target item's features as a "query" to filter the user's history. It dynamically assigns higher weights to past reviews and sentences that align with the current item's characteristics.
- Latent Feature Integration: To account for non-textual biases (e.g., a user's general rating tendency), the model incorporates ID-based embeddings via a Multi-Layer Perceptron (MLP).
- Prediction Layer: A Factorization Machine (FM) models the complex pairwise interactions between the final user and item embeddings to predict the rating.
Experimental Evaluation
The model was rigorously tested on ten datasets, including nine Amazon product categories and Yelp restaurant reviews. Experiments were conducted on both "5-core" and "10-core" subsets to evaluate performance across different data densities.
Performance Gains
AHN achieved state-of-the-art results across all benchmarks. The performance gap between AHN and existing models like DeepCoNN, D-ATT, and MPCN was particularly pronounced in denser datasets (10-core), where more textual information is available for the attention mechanism to exploit.
Key findings include:
- Contextual Filtering: The asymmetric approach effectively focuses on relevant past experiences, proving that treating users and items differently is crucial for accuracy.
- Noise Reduction: The hierarchical attention successfully filters out non-informative sentences common in reviews.
- Interpretability: By examining attention weights, the model can provide human-readable justifications for its recommendations, highlighting specific past reviews that influenced the current prediction.
Conclusion
The Asymmetrical Hierarchical Network addresses a critical oversight in review-based recommendation by acknowledging the asymmetry between user and item data. By leveraging item-guided attention, AHN provides both superior predictive accuracy and meaningful interpretability. Future work may explore the integration of multi-modal features, such as images, to further enhance the contextual filtering process.
Summary updated