NFARec: A Negative Feedback-Aware Recommender Model

Xinfeng Wang, Fumiyo Fukumoto, Jin Cui, Yoshimi Suzuki, Dongjin Yu

TL;DR

This paper introduces NFARec, a novel recommender system that effectively leverages negative feedback, which is often overlooked by existing models. By integrating sequential patterns with a Transformer Hawkes Process and structural patterns with a feedback-aware hypergraph convolution, NFARec significantly improves recommendation accuracy.

What Problem Does This Paper Address?

Recommender systems are crucial for navigating vast online content, but they often struggle with the nuances of user feedback. While positive interactions are readily incorporated, negative feedback (e.g., low ratings, dislikes) is frequently ignored or treated simplistically. This leads to several issues: (1) users might provide conflicting feedback on the same item, confusing models that treat all interactions equally; (2) even if an item is disliked, users might still interact with it, and models fail to capture this complex behavior. Existing Graph Neural Network (GNN) approaches, while good at capturing high-order collaborative signals, often neglect the valuable information embedded in negative feedback, leading to suboptimal user representations and recommendations.

Methods

NFARec tackles this problem through a two-pronged approach:

  • Sequential Representation Learning: This module uses the Transformer Hawkes Process (THP) to model user interaction sequences. The THP is particularly adept at capturing the influence of past events on current ones, with diminishing impact over time. Crucially, NFARec incorporates an auxiliary task: predicting the sentiment polarity (positive/negative) of the next interaction. This helps the model understand user preferences even when they interact with items they might not prefer, by learning the sentiment relatedness between sequential interactions.
  • Structural Representation Learning: This component employs a two-phase Hypergraph Convolution (HGC) approach.
  • Phase 1 (Interactive Hypergraph Convolution): This phase captures higher-order relationships beyond simple pairwise user-item interactions. A user hypergraph is constructed where a hyperedge connects a user to all items they've interacted with.
  • Phase 2 (Multi-order Feedback-guided Aggregation): This is the core innovation. It guides the HGC operator to efficiently propagate information between neighboring nodes by considering feedback. A feedback correlation matrix is built, and then iteratively used to construct multi-order feedback correlations. This allows the model to learn optimal message-passing paths, prioritizing interactions that align with user sentiment and avoiding those that conflict.

Datasets

The paper evaluates NFARec on five publicly available datasets with rating scores:

  • Yelp 2023
  • Beauty
  • Books
  • Recipes
  • MovieLens

For these datasets, interactions with ratings below 4 are treated as negative feedback, while others are positive. Users and items with fewer than a specified number of interactions are excluded.

Key Insights and Ideas

  • Negative feedback is a goldmine, not a nuisance: The core idea is that negative feedback provides crucial signals about user preferences and dislikes that are essential for accurate recommendations. Ignoring it leads to a significant performance gap.
  • Sequential sentiment matters: Predicting the sentiment of future interactions, not just the interaction itself, helps models understand user exploration and nuanced preferences.
  • Feedback guides graph convolutions: Instead of treating all connections in a graph equally, NFARec uses feedback to intelligently route information, making graph-based recommendations more context-aware.
  • Conflicting feedback is a key challenge: The paper highlights that users can have opposing views on similar items, and models need to disentangle these to avoid flawed representations.

Key Results

NFARec consistently outperforms state-of-the-art baselines across all five datasets. For instance, on the Yelp 2023 dataset, NFARec surpasses the runner-up XSimGCL by 4.2-5.6% in Recall@k and 4.1-10.9% in NDCG@k. Similar significant improvements are observed on other datasets, demonstrating the effectiveness of NFARec's negative feedback-aware approach. Ablation studies confirm that both sequential and structural components, especially the feedback-aware structural part, are vital for NFARec's superior performance.

Limitations and Future Work

  • Computational Complexity: The paper notes that NFARec's complexity is dominated by O(|I|²), where |I| is the number of items, making it computationally intensive for very large item sets.
  • Implicit Negative Feedback: The authors suggest exploring the impact of other forms of implicit negative feedback, such as skipping content, which were not explicitly modeled in this work.
  • Interpretability: Future work aims to incorporate large language models to enhance the interpretability of recommendations.
  • Dataset Sensitivity: While NFARec performs well, the paper acknowledges that the optimal order of feedback correlation can vary across datasets, suggesting that adaptive approaches might be beneficial.

Summary updated

BibTeX
@inproceedings{wang2024nfarec,
  author = {Xinfeng Wang and Fumiyo Fukumoto and Jin Cui and Yoshimi Suzuki and Dongjin Yu},
  title = {{NFARec: A Negative Feedback-Aware Recommender Model}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2024},
  doi = {10.1145/3626772.3657809},
}
Figure