[论文解读] DKN: Deep Knowledge-Aware Network for News Recommendation

Hongwei Wang, Fuzheng Zhang+2 more|arXiv (Cornell University)|Jan 25, 2018

Topic Modeling参考文献 50被引用 188

一句话总结

DKN 将知识图谱表示与多通道、词-实体对齐的 CNN 相结合，创建知识感知的新闻嵌入，并对用户历史进行注意力机制以预测新闻 CTR。它在 Bing News 数据集上优于最先进的基线，并且知识与注意力组件带来额外增益。

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.

研究动机与目标

• 通过在在线新闻中创建个性化推荐来缓解信息过载。
• 将外部知识图谱纳入以捕捉新闻项之间潜在的知识层级关联。
• 开发一个多通道 KCNN，使词向量与实体表示对齐以实现新闻理解。
• 通过对点击历史的注意力机制建模动态用户兴趣。
• 在真实的 Bing News 数据上展示相对于强基线的提升。

提出的方法

• 通过实体链接和一跳上下文扩展，为每条新闻构建知识图谱丰富的表示。
• 用 TransD 学习实体嵌入并推导上下文嵌入以丰富实体表示。
• 提出一个多通道的知识感知 CNN (KCNN)，其中词向量、实体向量和上下文嵌入形成对齐的通道。
• 用变换 g(E) 将词空间和实体空间对齐，实现多通道输入的联合卷积。
• 使用注意力网络对用户的点击历史相对于当前候选新闻进行加权。
• 在拼接用户与新闻嵌入后，使用下游神经网络预测 CTR。

实验结果

研究问题

• RQ1 如何将外部知识图谱整合到新闻推荐中，以捕捉文章之间潜在的知识层级关联？
• RQ2 在该领域中，词-实体对齐的 KCNN 是否能比传统的仅词 CNN 的表示更好？
• RQ3 注意力机制是否能有效从历史点击中建模出对当前候选新闻的多样化用户兴趣？
• RQ4 将知识图谱上下文引入与仅使用基础文本表示相比，在新闻 CTR 预测中能带来多大 empiricial 增益？

主要发现

模型	F1	AUC	p-value
DKN	68.9 ± 1.5	65.9 ± 1.2	-
LibFM	61.8 ± 2.1 (-10.3%)	59.7 ± 1.8 (-9.4%)	<10^{-3}
LibFM(-)	61.1 ± 1.9 (-11.3%)	58.9 ± 1.7 (-10.6%)	<10^{-3}
KPCNN	67.0 ± 1.6 (-2.8%)	64.2 ± 1.4 (-2.6%)	0.098
KPCNN(-)	65.8 ± 1.4 (-4.5%)	63.1 ± 1.5 (-4.2%)	0.036
DSSM	66.7 ± 1.8 (-3.2%)	63.6 ± 2.0 (-3.5%)	0.063
DSSM(-)	66.1 ± 1.6 (-4.1%)	63.2 ± 1.8 (-4.1%)	0.045
DeepWide	66.0 ± 1.2 (-4.2%)	63.3 ± 1.5 (-3.9%)	0.039
DeepWide(-)	63.7 ± 0.9 (-7.5%)	61.5 ± 1.1 (-6.7%)	0.004
DeepFM	63.8 ± 1.5 (-7.4%)	61.2 ± 2.3 (-7.1%)	0.014
DeepFM(-)	64.0 ± 1.9 (-7.1%)	61.1 ± 1.8 (-7.3%)	0.007
YouTubeNet	65.5 ± 1.2 (-4.9%)	63.0 ± 1.4 (-4.4%)	0.025
YouTubeNet(-)	65.1 ± 0.7 (-5.5%)	62.1 ± 1.3 (-5.8%)	0.011
DMF	57.2 ± 1.2 (-17.0%)	55.3 ± 1.0 (-16.1%)	<10^{-3}

• DKN 在 F1 和 AUC 上显著优于基线，报告的增益范围为 F1 2.8%~17.0%、AUC 2.6%~16.1%。
• 引入知识图谱及其上下文相对于无知识或无上下文的变体，能带来可度量的改进。