DKN: Deep Knowledge-Aware Network for News Recommendation (1801.08284v2)

Abstract: Online news recommender systems aim to address the information explosion of news and make personalized recommendation for users. In general, news language is highly condensed, full of knowledge entities and common sense. However, existing methods are unaware of such external knowledge and cannot fully discover latent knowledge-level connections among news. The recommended results for a user are consequently limited to simple patterns and cannot be extended reasonably. Moreover, news recommendation also faces the challenges of high time-sensitivity of news and dynamic diversity of users' interests. To solve the above problems, in this paper, we propose a deep knowledge-aware network (DKN) that incorporates knowledge graph representation into news recommendation. DKN is a content-based deep recommendation framework for click-through rate prediction. The key component of DKN is a multi-channel and word-entity-aligned knowledge-aware convolutional neural network (KCNN) that fuses semantic-level and knowledge-level representations of news. KCNN treats words and entities as multiple channels, and explicitly keeps their alignment relationship during convolution. In addition, to address users' diverse interests, we also design an attention module in DKN to dynamically aggregate a user's history with respect to current candidate news. Through extensive experiments on a real online news platform, we demonstrate that DKN achieves substantial gains over state-of-the-art deep recommendation models. We also validate the efficacy of the usage of knowledge in DKN.

Deep Knowledge-Aware Network for News Recommendation: An Overview

The paper presents DKN (Deep Knowledge-Aware Network), an innovative model addressing the challenge of news recommendation by integrating knowledge graph representation into deep learning. This model leverages external knowledge to enhance the semantic understanding and prediction capabilities, particularly focusing on content-based click-through rate (CTR) prediction.

Key Contributions

The authors highlight three pivotal challenges in news recommendation:

1. Time-Sensitivity of News: Unlike static items such as movies or products, news articles quickly lose relevance as they are frequently replaced by newer content.
2. Diverse User Interests: Users generally have multifaceted interests, making it necessary to dynamically measure user preferences based on their historical interactions.
3. Condensed Language with Knowledge Entities: News titles are typically concise and rich in knowledge entities, relationships, and common sense, which traditional semantic models fail to adequately capture.

To address these challenges, the authors propose the DKN framework, characterized by the integration of knowledge graph representations and a novel convolutional neural network architecture.

DKN Framework

DKN's architecture consists of several key components:

• Knowledge-Aware Convolutional Neural Network (KCNN): This is the core of the DKN framework, designed to fuse semantic-level and knowledge-level representations. KCNN employs a multi-channel approach where words and entities are treated as separate channels. The alignment of word and entity embeddings during convolution ensures these different but complementary sources of information are effectively combined.
• Attention Module: To cater to users’ diverse interests, an attention mechanism dynamically weighs the user's click history relative to the current candidate news. This way, the model aggregates historical preferences more precisely.

Empirical Results

Through extensive experiments using a real-world dataset from Bing News, DKN demonstrates significant improvements over several state-of-the-art methods including LibFM, KPCNN, DSSM, DeepWide, DeepFM, YouTubeNet, and DMF. The key performance metrics, F1 and AUC, are improved by ranges of 2.8% to 17.0% for F1 and 2.6% to 16.1% for AUC, highlighting the efficacy of incorporating knowledge entities and user-specific attention mechanisms.

Model Variants and Sensitivity Analysis

The authors explore various configurations, demonstrating that including contextual entity embeddings enhances performance and that using TransD for knowledge graph embedding yields superior results compared to other methods like TransE or TransH. The results reveal that both word and entity embeddings must be appropriately transformed and that non-linear mapping outperforms linear mapping.

Practical and Theoretical Implications

• Practical: By integrating knowledge representation, DKN is well-suited for dynamic and context-rich environments like online news platforms. Its architecture enhances personalized recommendations, enabling platforms to push relevant content more effectively, thus improving user engagement.
• Theoretical: This work bridges semantic representation with symbolic knowledge, advancing the understanding of how disparate sources of information can be harmonized for improved prediction tasks. It paves the way for further research into knowledge-integrated deep learning models for varied applications.

Future Directions

Further research might investigate the application of DKN principles to other domains where temporal dynamics and rich contextual dependencies are critical, such as social media, real-time advertising, or personalized education. Enhanced interpretability and further optimization of attention mechanisms could yield even more robust performance enhancements across different user behaviors and content types.

In conclusion, the DKN framework significantly advances the field of personalized news recommendation by effectively integrating deep learning with knowledge graph representations, addressing critical challenges in time-sensitivity, diversified user interests, and complex language structures.