基於圖卷積網路探討推薦系統中使用者之共同興趣

Abstract

推薦系統對現代社會產生了重大影響，從根本上改變了人們的消費習慣，而分析用戶行為是推薦系統研究領域中最熱門的議題之一，利用用戶的歷史互動資料來了解、分析用戶的行為，可以提升推薦系統的準確度。以往的研究主要集中在分析單個用戶的行為，獲取更多隱含訊息，或者尋找有相似行為的用戶來提高推薦的表現。然而，人們傾向於一起做出購買決定，例如，當和朋友去看電影時，我們通常會根據我們的共同興趣來決定看哪部電影。為了讓推薦系統有更進一步的正確率，找到使用者的共同興趣將是一個關鍵的問題。一個簡單找到使用者共同興趣的方法是直接取每個用戶行為的交集，然而這樣每個用戶行為中的潛在行為將被忽略，且共同興趣的可解釋性將受到限制。 為了更好地模擬每個用戶與項目交互和共同興趣的行為，我們提出了共同興趣（COIN）模型來尋找這項工作中的共同興趣。COIN 模型利用 Graph Convolution Network 的特性對 user-item 交互圖中的高階特徵（high-order) 進行建模，並利用 item 標籤屬性進行解釋。最後，在三個真實世界數據集上的大量實驗結果顯示了我們的 COIN 模型與其他方法相比的有效性。

Recommender systems have a significant impact on human society, and it has changed people’s consumption habits. Analyzing user behavior is one of the most popular topics among the recommendation research field. Previous studies focused on either analyzing the behavior of a single user, obtaining more implicit information, or finding similar users to enhance recommendation performance. However, people tend to make purchase decisions together, e.g., when going to the movies with friends, we usually decide which movie to watch based on our common interests. In order to have further improvement in recommendation, finding common interests will be a critical problem. A simple way to find common interests is to directly take the intersection of each user’s behavior. However, the latent semantics within the behavior of each user will be ignored, and the explainability of common interests will be limited. To better model the behavior of each user with item interaction and common interest, we propose the COmmon INterest (COIN) model for finding the common interests in this work. The COIN model leverages the characteristics of Graph Convolution Network to model high-order features in the user-item interaction graph and make use of the item tag attribute for the explanation. Lastly, extensive experimental results on three real-world datasets show the effectiveness of our COIN model compared with baseline methods.