- 學位論文
運用生成式對抗網路緩解推薦系統冷啟動使用者問題
Mitigating New User Cold-Start Problem in Recommendation Systems with Generative Adversarial Networks
摘要
隨著網際網路快速的發展，促進網路服務的多元開發，也讓現代人的生活越來越依賴網路服務平台，其中，對網路服務平台供應商而言，是否能獲得新進使用者的青睞是影響收益的關鍵因素。由於傳統的推薦系統演算法是基於使用者行為來進行產品推薦，在面對新進使用者(new users)時會因為只有他們少許的行為資料，推薦演算法難以從中分析出使用者興趣，而此問題被稱為「新進使用者冷啟動問題」(new user cold-start problem)，這樣的情境使得推薦系統難以了解新進使用者的興趣，更降低了推薦的效果。過去的研究試圖使用額外的資訊來解決新進使用者冷啟動問題，例如：使用者的性別、職業或社群網路資訊等等，但是資訊隱私與個人資訊保護使得先前的研究無法應用。 在本篇論文中，我們提出了一個端到端(end-to-end)基於生成式對抗網路(GAN)的推薦演算法來緩解使用者冷啟動的問題，此方法不需使用到額外的資訊，此方法由兩個神經網路組成：生成器與判別器，利用豐富的使用者資訊來訓練生成器，而生成器會學習模擬新進使用者變成豐富使用者的評分分佈，同時，判別器會負責分辨生成器模擬出的資訊與真實資訊，最後訓練完成的生成器就可以替冷啟動使用者進行推薦。此外，我們設計了「返老還童」機制來將豐富使用者還原到他剛加入平台時的冷啟動狀態，根據在三個不同領域資料集的實驗結果，我們提出的方法都遠遠優於其他冷啟動推薦演算法。
並列摘要
Recommending appropriate items to new users creates a good sense of belonging that helps e-services retain newcomers. Such recommendations are difficult because there are limited behavior data (e.g., item ratings) for recommendation systems to infer user preferences. Research on the new user cold-start recommendation problem generally leverages side information (e.g., demographic data) to suggest items to new users. This approach, however, is impractical due to privacy concerns. In this paper, we propose an end-to-end GAN-based recommendation system that makes no use of side information to resolve the new user cold-start recommendation problem. To the best of our knowledge, this is the first end-to-end GAN-based method that focuses on the new user cold-start recommendation. The proposed model explores the merit of GAN that enables perfect data generation given imperfect data, and consists of two adversarial networks: a generative network and a discriminative network. The generative network learns to predict item ratings that cold-start users would make in the future given their limited rating behavior data. The predicted ratings are evaluated by the discriminative network trained for determining whether the ratings are perfect enough. Moreover, a novel rejuvenation function and a relevant item loss are incorporated into proposed model to enhance the performance of the predictions make by the learned generative network. Experiments based on three real-world datasets across domains of movies, books, and gift cards demonstrate that the designed rejuvenation function and the relevant item loss are effective in guiding our generative network to infer item ratings of cold-start users. Also, our proposed model significantly outperforms many well-known recommendation systems, thus indicating that the suggested items are relevant to the preferences of cold-start new users.