Introduction
Personalized news recommendation is a critical component of modern digital content consumption. However, traditional approaches often rely on a single, "one-size-fits-all" embedding to represent a user's diverse interests. HieRec (Hierarchical User Interest Modeling) addresses this limitation by modeling user interests as a three-level hierarchy—subtopic, topic, and overall interest—to better capture the multi-grained nature of what people actually like to read.
The Problem: The Multi-Grained Reality of Interest
Most existing systems compress a user's entire reading history into a single mathematical vector. This approach struggles with the "multi-grained" reality of human interest. For instance, a user might have a broad interest in Sports but a very specific interest in Football, while having no interest in Golf. A single vector often blurs these distinctions, resulting in profiles that are either too generic or too narrow. Furthermore, different articles require different levels of matching; a major stadium announcement might appeal to a general sports fan, whereas a technical analysis of a formation only appeals to a subtopic specialist.
The HieRec Framework
HieRec introduces a two-pronged approach to hierarchical modeling:
Hierarchical User Interest Modeling
The system constructs a user profile across three levels of granularity:
- Subtopic-level: Clicked news items are grouped by subtopic (e.g., "Football"). An attention mechanism highlights the most informative clicks, which are then combined with subtopic embeddings.
- Topic-level: Subtopic interests are aggregated into broader categories (e.g., "Sports"). This level uses "topic-level attention" that accounts for both content and click frequency.
- User-level: Finally, topic-level representations are aggregated into a global embedding representing the user’s overall interest profile.
Hierarchical User Interest Matching
Instead of a single similarity score, HieRec calculates three separate scores ($o_s, o_t, o_g$) by matching a candidate article against the user's subtopic, topic, and global interest levels. The final recommendation score is a weighted combination of these, allowing the model to target the user at the most appropriate level of granularity for a specific piece of news.
Experimental Evaluation
The model was evaluated on two large-scale datasets: the public MIND (Microsoft News Dataset) and a private Feeds dataset from a commercial news app. HieRec consistently outperformed top-tier baselines such as NRMS, NAML, and LSTUR. On the MIND dataset, HieRec achieved an AUC of 67.95, significantly surpassing the best baseline (KRED at 65.89).
Beyond accuracy, HieRec showed a clear advantage in diversity, measured by Intra-List Average Distance (ILAD). By recognizing the breadth of a user's interest tree, the system helps alleviate the "filter bubble" problem, providing a more varied and representative feed.
The Importance of Hierarchy
Ablation studies confirmed that the multi-level hierarchy is essential for performance. Removing either the subtopic or topic levels led to a measurable drop in both AUC and nDCG metrics, proving that multi-grained modeling is the key driver of HieRec's success.
Conclusion and Future Directions
HieRec demonstrates that modeling user interests as a hierarchy significantly improves both the accuracy and diversity of news recommendations. While the model currently relies on high-quality topic labels, it remains robust even when fine-grained data is limited. Future work includes exploring Federated Learning to address privacy concerns associated with detailed user interest modeling and improving performance for "cold start" users who lack a developed interest hierarchy.
Summary updated