Abstract
Personalized news recommendation remains a challenge due to the coarse-grained nature of traditional user profiles. This paper introduces a novel approach that constructs high-fidelity profiles by extracting (entity, aspect) pairs from user comments. By capturing the specific context of a user's interest (e.g., "Tunisia" in the context of "Tourism" vs. "Elections"), the proposed system significantly enhances recommendation accuracy. Experimental results show a 10% improvement in precision over entity-centric baselines, demonstrating the value of leveraging rich, subjective data from comment sections.
Introduction
In an era of information overload, news platforms must effectively filter content to maintain user engagement. Most existing recommendation engines rely on broad categories (e.g., "Politics") or simple entity tags (e.g., "Tunisia"). However, these methods fail to capture the nuanced interests of readers who may care about an entity in one context but not another.
Furthermore, many systems utilize short-form data like social media posts, which often lack the depth required for detailed profiling. This research addresses these gaps by utilizing long-form news comments to build fine-grained user profiles based on (entity, aspect) tuples, treating both users and articles as collections of these specific contextual markers.
Methodology
The core of the proposed system is a content-filtering strategy that models interests through entity-aspect conjunctions.
Profile Extraction
The extraction process involves two primary components:
- Entities: Using OpenNLP for segmentation and OpenCalais for identification, the system extracts well-defined concepts such as persons, locations, and organizations.
- Aspects: Using Zemanta, the system identifies attributes or abstract objects related to those entities (e.g., "Work permit" associated with an immigration-related entity).
Recommendation Logic
When a user interacts with a "seed" article, the system generates recommendations through a two-step ranking process:
- Contextual Relevance ($L_1$): Articles from the same timeframe are ranked based on their similarity to the current seed article's profile.
- Historical Alignment ($L_2$): The same articles are ranked based on their similarity to the user's long-term historical profile.
- Aggregation: These lists are merged to provide recommendations that are both immediately relevant and aligned with enduring preferences.
Similarity is calculated using a modified Cosine Similarity metric, where vector dimensions represent the frequency of specific (entity, aspect) tuples.
Dataset and Evaluation
The study utilized a dataset crawled from four major news platforms: The Independent, The Telegraph, CNN, and Al-Jazeera. The dataset comprises nearly half a million comments across over 50,000 articles.
| Metric | Count |
|---|---|
| Total Comments | 482,073 |
| Independent Articles | 26,096 |
| Telegraph Articles | 23,154 |
| CNN Articles | 535 |
| Al-Jazeera Articles | 303 |
Evaluation was conducted using 189 recommendation rounds for 23 active users. As illustrated in Figure 1, the dataset exhibits a strong bias toward political news across all subsets (seed articles, commented articles, and published articles).
Results and Discussion
The experimental results validate the hypothesis that combining entities and aspects provides a superior signal for recommendation compared to using either in isolation.
- Superior Precision: The "Global Profile" (entity + aspect) achieved a P@5 of 0.52, outperforming the entity-only baseline (0.412) by approximately 10%.
- High Ranking Quality: An NDCG@5 of 0.855 suggests the system is highly effective at surfacing the most relevant content at the top of the recommendation list.
- Contextual Necessity: Models using only aspects performed the worst (P@5 of 0.396), confirming that abstract aspects must be anchored to specific entities to be meaningful.
A key insight is that news comments serve as a "goldmine" of data. Unlike shorter social media posts, comments provide sufficient linguistic context to extract stable (entity, aspect) pairs. Additionally, the system employs dynamic updates (triggered every 100 comments) to ensure profiles evolve alongside user interests.
Limitations and Future Work
While promising, the system faces several challenges. The current evaluation relies on comments as the sole indicator of interest, likely underestimating true precision since users often read without commenting. The "cold start" problem also remains, as the system requires a threshold of 100 comments to build a reliable profile.
Future research will focus on diversification techniques to prevent "echo chambers" by recommending articles that lie slightly outside a user's established profile, thereby broadening their news consumption while maintaining relevance.
Summary updated