用於推薦系統的基於模擬輔助信息路徑的圖嵌入模型

實時的個性化推薦系統對於在線購物平台來說至關重要，它可以有效提高用戶購買率。現有的推薦模型在模型訓練的時候嚴重依賴於大量密集的數據。但事實上，用戶的數據量可能很會少並且很稀疏，尤其是對於某些初創公司和新平台而言。此外，用戶興趣是會隨時間而變化的。一個高效的推薦系統應該可以實時地捕捉到用戶偏好的更新。為了解決這些挑戰，我們提出使用異構網絡去對用戶和物品之間的關係進行建模，並開發一種稱為SSIP的圖嵌入模型去學習物品的嵌入表達。SSIP根據用戶的歷史行為和商品的輔助信息來模擬用戶的訂購行為，從而解決初創公司數據量少和稀疏的問題。在我們的模型中，用戶的嵌入表達是根據他們最近瀏覽商品的嵌入表達生成的，因此可以被更有效率地進行更新。我們進行了大規模的離線實驗和在線A/B測試來衡量我們的模型。離線實驗表明，SSIP的效果優於經典推薦模型和最新的圖嵌入模型，在真實數據集上表現的F1分數提高了23.4%。 KKday是一家賣旅遊相關產品的電子商務公司，我們在KKday上進行的在線A/B測試也取得了CTR提升15.14%的成效，證明了我們方法的有效性。

關鍵字

網路嵌入；推薦系統；電子商務推薦

並列摘要

A real-time personalized recommendation system is of vital importance for online shopping platforms, as it could effectively increase the purchase rate. Existing recommendation models rely heavily on a huge amount of dense data for model training. However, in reality, the amount of user data could be small and the data could be sparse, especially for some startup companies and new platforms. Furthermore, user interest is dynamic over time. An efficient recommendation system should update user preferences in real time. To tackle the challenges, we propose using heterogeneous networks to model the relationships between users and items, and have developed a graph embedding model called SSIP to learn the embedding for items. SSIP simulates users' ordering behavior based on their historical behavior and the side information of items, by which we can address the small amounts of data and data sparsity issues in startup companies. In our model, user embeddings are generated from embeddings of their recent browsed items, so that they can be updated efficiently. We conduct extensive offline experiments and online A/B tests to evaluate our model. The offline experiments show that SSIP outperforms the classical recommendation models and the state-of-the-art graph embedding models and on the real-world dataset (with an improvement of 23.4% on F1 score). Online A/B test results in KKday, which is a B2C e-commerce company for travel products, also demonstrate the effectiveness of the proposed method (with an improvement of 15.14% on CTR).

並列關鍵字

network embedding ； recommendation system ； e-commerce recommendation

參考文獻

[1] M. Balabanović and Y. Shoham. “Fab: Content-based, collaborative recommendation”. In: (1997).

Google Scholar

[2] Xiangnan He et al. “Fast Matrix Factorization for Online Recommendation with Implicit Feedback”. In: Proceedings of the 39th International ACM SIGIR Conference