SentiEntityRec: Entity-level sentiment perception graph neural network for news recommendation

Qingshuai Wang, Jiahao Wang, Noor Farizah IbrahimJournal of King Saud University - Computer and Information Sciences 2025 · DOI

This paper, published in the Journal of King Saud University - Computer and Information Sciences, introduces SentiEntityRec, a framework designed to make news recommendations more "emotionally intelligent" by looking at sentiment at the entity level rather than the article level.

TL;DR

SentiEntityRec is a novel graph neural network (GNN) framework that improves news recommendations by incorporating fine-grained, entity-level sentiment vectors. By distinguishing how a user feels about specific companies or people mentioned in a story—rather than just the overall topic—the model achieves state-of-the-art performance on the industry-standard MIND dataset.

What Problem Does This Paper Address?

Most modern news recommendation systems rely on two things: semantics (what is the story about?) and user history (what did the user click before?). While effective, these systems often suffer from "sentiment blindness."

The authors point out a critical flaw in document-level sentiment analysis: a single news article often contains conflicting sentiments. For example, a headline reading "Alibaba reports record sales while Tencent faces regulatory challenges" would likely be labeled "neutral" by a standard model. However, an investor in Alibaba would find this news positive, while a Tencent stakeholder would see it as negative.

Existing Graph Neural Networks (GNNs) capture complex relationships between users and news, but they typically treat entities (like "Alibaba" or "Tencent") as static semantic nodes. They ignore the "sentimental resonance" that actually drives a user to click or ignore a story.

Methods

The SentiEntityRec architecture consists of several sophisticated layers designed to fuse text, relationships, and emotions:

  1. Sentiment Perception Component (SPC): This is the "eyes" of the model. It uses a pre-trained model (NewsMTSC) to extract entities from news titles and abstracts and assigns them a 3-dimensional sentiment probability vector (Negative, Neutral, Positive).
  2. Dual-Entity Encoding: The model creates two separate representations for every entity:
  3. Semantic Encoding: Uses TransE embeddings to capture the entity's place in a knowledge graph (e.g., knowing that "Alibaba" is a "Company").
  4. Sentiment-Aware Encoding: Fuses the semantic embedding with the 3D sentiment vector from the SPC.
  5. Graph Neural Network (GNN) Layer: The authors build a global news graph and a semantic entity graph based on user click sequences. They utilize a Gated Graph Neural Network (GGNN) to propagate information across these graphs, which helps the model learn the sequential "flow" of a user's reading habits.
  6. Attention-Based Aggregation: Finally, an attention mechanism weighs the importance of the news content, the general entity info, and the sentiment-specific info to create a final "interest profile" for the user and a "relevance score" for candidate news articles.

Datasets

The researchers evaluated their model using the MIND (Microsoft News Dataset), the gold standard for news recommendation research.

  • MIND-small: 50,000 users and their behavior logs.
  • MIND-large: 1,000,000 users and over 24 million clicks. The use of both versions ensures that the model's performance is consistent across different scales of data.

Key Insights and Ideas

  • The Nuance Gap: The authors found that only 47.7% of entities share the same sentiment polarity as the overall document they appear in. This confirms that document-level sentiment is a noisy, unreliable signal for personalization.
  • Sentiment as a Feature, Not Just a Label: Instead of just classifying a news item as "happy" or "sad," the model treats sentiment as a continuous vector that enriches the entity's identity. This allows the GNN to find "sentimental clusters" in user behavior.
  • GGNN Superiority: In their tests, Gated Graph Neural Networks (GGNN) outperformed standard Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT). This suggests that the order in which users read news (sequential behavior) is vital when combined with sentiment data.

Key Results

SentiEntityRec outperformed 16 baseline models, including high-performing recent frameworks like GLORY and GLoCIM.

  • Performance Boost: On the MIND-large dataset, SentiEntityRec achieved an AUC of 69.46, a significant improvement over the previous best baseline (GLoCIM at 69.06).
  • Ablation Success: When the authors removed the sentiment component ($e^{emo}$), the model's performance dropped across all metrics (AUC, MRR, nDCG). This proves that the sentiment data—not just the GNN structure—is doing the heavy lifting.
  • Optimal Hyperparameters: The model performs best when looking at the 8 nearest news neighbors and 10 adjacent entities, suggesting that "local" context is more important than broad, global graph connections.

Limitations and Future Work

While powerful, the authors acknowledge a few hurdles:

  • Computational Complexity: GNNs are computationally expensive. Training on 1,000 samples with 3 "hops" in the graph takes significantly longer and uses more GPU memory than simpler models, which could make real-time deployment difficult.
  • Entity Disambiguation: The current model may struggle with synonyms or related words that refer to the same entity but aren't explicitly linked in the knowledge graph.
  • Future Directions: The team plans to focus on GNN scalability to make the system faster and more efficient for production environments, as well as improving the model's ability to recognize linguistic synonyms.

Summary updated

Abstract

Precise news recommendations are critical in today’s digital landscape. However, conventional approaches overlook fine-grained sentiment nuances associated with individual entities in news content. This paper presents SentiEntityRec, a novel graph neural network framework that enriches traditional entity embeddings using a global graph-enhanced model by incorporating sentiment vectors. Experiments conducted with MIND datasets have shown that SentiEntityRec surpasses existing models in key performance metrics. For example, the proposed model improves +0.5% AUC over GLoCIM. These results underscore the superior efficacy of incorporating entity-sentiment analysis into graph-based news recommendation systems.
BibTeX
@article{wang2025sentientityrec,
  author = {Qingshuai Wang and Jiahao Wang and Noor Farizah Ibrahim},
  title = {{SentiEntityRec: Entity-level sentiment perception graph neural network for news recommendation}},
  journal = {Journal of King Saud University - Computer and Information Sciences},
  year = {2025},
  doi = {10.1007/s44443-025-00087-2},
}
Figure