近年來，將知識圖譜應用於推薦系統模型獲得了不錯的成效，其中重要的模型如RippleNet和CKAN等。這些模型利用知識圖譜來擴展實資訊，其原理類似水波擴散而形成了漣漪，擴展的資訊可強化「使用者偏好」與「推薦項目」之間的關聯度。然而，在RippleNet和CKAN的模型中，「使用者偏好」與「推薦項目」的特徵學習網路高度耦合，這可能會降低候選項目之間的鑑別力，並導致多層擴散無法再提升效能。另一方面，知識圖譜擴散時，擴增實體數會隨著擴散層級數呈指數成長。故此，前人使用了隨機採樣對三元組做刪減，以降低計算量。然而隨機採樣可能會使嵌入特徵的學習限於固定的三元組而缺乏知識涵蓋度。為解決上述問題，本研究試圖提出解耦合的網路架構，並對知識圖譜傳播過程的採樣方式進行探討。
首先，我們提出了解耦合的知識傳播網路架構，藉由分離「使用者偏好」與「推薦項目」網路模型的耦合性，讓模型更能夠區別使用者偏好與推薦項目，以增進候選推薦項目之間的鑑別度。基於此架構，我們在 MovieLens 1M推薦系統資料集與 Microsoft Satori知識圖譜上進行實驗，結果顯示AUC可達 0.9299，優於RippleNet與CKAN，並且效能可隨著更多層級而持續增加。進一步，我們針對採樣方式提出了三種改進策略：平衡採樣、非重複採樣和動態採樣。平衡採樣是固定種子與維持三元組數目的平衡，以避免採樣有所偏頗；非重複採樣是不去擴展重複的實體；動態採樣則是在每訓練達到一定次數後就重新採樣。我們混合三種採樣策略，在解耦合知識傳播網路進行實驗，AUC最高可達到0.9358。最後，我們對公開資源WikiData的知識圖譜進行實驗，驗證此模型對不同知識圖譜也可展現優越的效能。

State-of-the-art knowledge-graph-based models, such as RippleNet and CKAN, have been successfully applied to recommendation systems. These models make use of knowledge graph to expand the entity information, which is similar to ripples, so as to strengthen the correlation between the user preferences and the candidate item. In RippleNet and CKAN, however, the networks of representation learning for "user preference" and "candidate item" are mutually couple, which might degrade the discriminability of the candidate items and limit the recommendation performance, especially for multi-levels propagation. On the other hand, when the entities are expanded through knowledge graph, the amount of entities increases exponentially with the number of diffusion levels. In earlier research, random sampling was utilized to restrict the set of triples, which however limits the scope of the expanded information used for learning as well as the prediction performance.
To tackle these issues, we first propose decoupled knowledge propagation network (DKPN) that decoupled the user preference and the candidate item so as to better distinguish the candidate items. A series of experiments were conducted on the MovieLens 1M recommendation dataset and the Microsoft Satori knowledge graph, and the results show AUC of DKPN can reach 0.9299, which is superior to RippleNet and CKAN, and the performance increases with the number of levels persistently. In addition, three sampling strategies were proposed to improve the model, including balanced sampling, non-duplication, and dynamic sampling. Experiments show these sampling strategies are effective and the highest AUC, 0.9358, can be obtained. Finally, DKPN was tested on a movie-related knowledge graph collected from the WikiData, and verified to achieve the best performance.

[1] A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, and O. Yakhnenko, “Translating embeddings for modeling multi-relational data,” in NIPS, 2013, pp. 2787–2795.
[2] Maximilian Nickel, Volker Tresp, Hans-Peter Kriegel Authors Info & Claims. “A Three-Way Model for Collective Learning on Multi-Relational Data” ICML Pages 809–816
[3] Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, Sebastian Riedel. ” Convolutional 2D Knowledge Graph Embeddings” arXiv:1707.01476 .
[4] D. Q. Nguyen, T. D. Nguyen, D. Q. Nguyen, and D. Phung, “A novel embedding model for knowledge base completion based on convolutional neural network,” in NAACL, 2018, pp. 327–333.
[5] Hongwei Wang, Fuzheng Zhang, Xing Xie, and Minyi Guo. 2018. “DKN: Deep Knowledge-Aware Network for News Recommendation”. In Proceedings of the 2018 World Wide Web Conference on World Wide Web. International World Wide Web Conferences Steering Committee, 1835–1844.
[6] G. Ji, S. He, L. Xu, K. Liu, and J. Zhao, “Knowledge graph embedding via dynamic mapping matrix,” in Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2015, pp. 687–696.
[7] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin. “Attention Is All You Need.”, 2017, arXiv: 1706.03762.
[8] X. Wang, D. Wang, C. Xu, X. He, Y. Cao, and T.-S. Chua, 2019 “Explainable reasoning over knowledge graphs for recommendation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, pp. 5329–5336.
[9] Sepp Hochreiter and Jürgen Schmidhuber. ”Long Short-Term memory. Neural computation”, 9(8):1735–1780, 1997.
[10] H. Wang, F. Zhang, J. Wang, M. Zhao, W. Li, X. Xie, and M. Guo, “Ripplenet: Propagating user preferences on the knowledge graph for recommender systems,” in Proceedings of the 27th ACM International Conference on Information and Knowledge Management. ACM, 2018, pp. 417–426.
[11] Ze Wang, Guangyan Lin, Huobin Tan, Qinghong Chen, and Xiyang Liu. “CKAN: Collaborative Knowledge-aware Attentive Network for Recommender Systems”, SIGIR 2020.
[12] Machine Learning, Kohavi, R. and Provost, F., “Machine Learning—Special Issue on Applications of Machine Learning and the Knowledge Discovery Process”, 30, 271-274.
[13] Wen Zhu1 , Nancy Zeng2, Ning Wang2S, “Ensitivity, Specificity, Accuracy, Associated Confidence Interval and ROC Analysis with Practical SAS® Implementations” , NESUG 2010.
[14] Andrew P.Bradley, “The use of the area under the ROC curve in the evaluation of machine learning algorithms, Pattern Recognition”, 30, 1145-1159.
[15] T. Hofmann, "Latent semantic models for collaborative filtering," ACM Transactions on Information Systems (TOIS), vol. 22, no. 1, pp. 89-115, 2004.
[16] T. Hofmann and J. Puzicha, "Latent class models for collaborative filtering," in IJCAI, vol. 99, no.1999, 1999.
[17] Y. Koren, R. Bell, and C. Volinsky, "Matrix factorization techniques for recommender systems," Computer, no. 8, pp. 30-37, 2009.
[18] Mikolov, Tomas, Chen, Kai, Corrado, Greg, and Dean, Jeffrey. “Efficient estimation of word representations in vector space”. 2013a, arXiv preprint arXiv:1301.3781.
[19] F. Manola, E. Miller, B. McBride, “RDF Primer, W3C Rec”. 10 February 2004.
[20] “E. Prud’hommeaux and A. Seaborne. SPARQL Query Language for RDF”. W3C Candidate Rec. 6 April 2006. http://www.w3.org/TR/rdf-sparql-query/
[21] Denny Vrandečić, Markus Krötzsch, “Wikidata: a free collaborative knowledgebase”. October 2014 pp 78–85