[Paper Review] A Zero Attentive Relevance Matching Networkfor Review Modeling in Recommendation System

Hansi Zeng, Zhichao Xu+1 more|arXiv (Cornell University)|Jan 16, 2021

Recommender Systems and Techniques40 references4 citations

TL;DR

This paper proposes ZARM, a novel interaction-based recommendation model that uses a relevance matching network with regularized training to dynamically extract user representations from reviews based on relevance to the target item, while employing a zero-attention mechanism to degrade gracefully when no relevant reviews exist. The model outperforms state-of-the-art siamese and interaction-based models on four Amazon benchmark datasets by effectively capturing user-item relevance and improving rating prediction accuracy.

ABSTRACT

User and item reviews are valuable for the construction of recommender systems. In general, existing review-based methods for recommendation can be broadly categorized into two groups: the siamese models that build static user and item representations from their reviews respectively, and the interaction-based models that encode user and item dynamically according to the similarity or relationships of their reviews. Although the interaction-based models have more model capacity and fit human purchasing behavior better, several problematic model designs and assumptions of the existing interaction-based models lead to its suboptimal performance compared to existing siamese models. In this paper, we identify three problems of the existing interaction-based recommendation models and propose a couple of solutions as well as a new interaction-based model to incorporate review data for rating prediction. Our model implements a relevance matching model with regularized training losses to discover user relevant information from long item reviews, and it also adapts a zero attention strategy to dynamically balance the item-dependent and item-independent information extracted from user reviews. Empirical experiments and case studies on Amazon Product Benchmark datasets show that our model can extract effective and interpretable user/item representations from their reviews and outperforms multiple types of state-of-the-art review-based recommendation models.

Motivation & Objective

• To address the suboptimal performance of existing interaction-based recommendation models despite their higher capacity and better alignment with human behavior.
• To overcome three key limitations in current interaction-based models: dilution of similarity signals in long reviews, overfitting due to limited user reviews, and reliance on textual similarity rather than true user opinion.
• To develop a model that can dynamically learn user representations based on relevance to the target item, while maintaining robustness when no relevant reviews are present.
• To improve the interpretability and effectiveness of review-based representations by using ground-truth user reviews as positive pairs for training the relevance matching module.

Proposed method

• The model employs a relevance matching module that treats each user review as a query to retrieve relevant content from the target item’s reviews, enabling discovery of semantically relevant information even in long text.
• A regularized auxiliary loss is introduced using the ground-truth user review and corresponding item review as a positive 'query-document' pair, improving the training of the relevance matching function.
• A zero-attention network is used to dynamically balance item-dependent and item-independent information from user reviews, allowing the model to degrade to a siamese model when no relevant reviews are found.
• Static user and item embeddings are learned via a multi-layer convolutional self-attention network to capture hierarchical representations from words to reviews.
• Final user representations are formed by combining dynamic representations from the interaction module and static embeddings, ensuring robustness in low-relevance scenarios.
• The model integrates the relevance matching module as an auxiliary loss in the training objective, enhancing the alignment between relevance scores and true user opinions.

Experimental results

Research questions

• RQ1 How can interaction-based models better capture user relevance to target items without relying on raw textual similarity?
• RQ2 What impact does using ground-truth user reviews as positive pairs have on training the relevance matching function?
• RQ3 How can a model maintain performance when no user reviews are relevant to the target item?
• RQ4 To what extent does the zero-attention mechanism improve robustness and generalization in low-relevance scenarios?
• RQ5 How do the relevance matching scores correlate with actual user opinions and item characteristics?

Key findings

• The ZARM model outperforms multiple state-of-the-art siamese and interaction-based models on four Amazon benchmark datasets in rating prediction accuracy.
• Approximately 50% of user-item pairs in the datasets have no relevant user reviews to the target item, with the highest proportion (60%) in the Video Games dataset.
• In the Toys & Games dataset, 15% of user-item pairs have more than three relevant reviews, indicating consistent user interests across similar items.
• Case studies show that high relevance scores correlate with shared keywords and semantic alignment between user reviews and item descriptions, such as 'Gyro Hercules' and 'durable'.
• The zero-attention mechanism successfully degrades to a siamese model when no relevant reviews exist, preserving performance in sparse scenarios.
• The relevance matching module with auxiliary loss significantly improves the model’s ability to identify semantically relevant reviews that reflect true user opinions on target items.