SentiGraphRec: sentiment perception entity graph for personalized news recommendation

Abstract

In the digital era, accurate news recommendation is essential for improving user engagement and supporting effective information delivery. Most existing methods mainly focus on semantic or structural signals, while largely overlooking users’ sentiment preferences toward entities in their reading histories. As a result, these approaches struggle to model dynamic sentiment patterns and fail to provide sufficiently personalized recommendations. To address this issue, we propose SentiGraphRec, a sentiment-aware entity graph model for personalized news recommendation. SentiGraphRec constructs a global entity graph that jointly encodes semantic relations, structural connections, and entity-level sentiment information. By explicitly modeling the interactions among users, news articles, and entities with sentiment orientations, the proposed framework captures both semantic relevance and affective preferences in a unified manner. Extensive experiments on public benchmark datasets demonstrate that SentiGraphRec consistently outperforms strong state-of-the-art baselines across multiple evaluation metrics. In particular, the model achieves notable improvements in recommendation accuracy and diversity. These results validate the effectiveness of sentiment-aware entity modeling and highlight the general applicability of SentiGraphRec for personalized news recommendation.