推薦系統在深度學習領域取得了顯著的成功，特別是對於具有大量互動記 錄的項目。然而，這些系統常常面臨由於項目屬性(如價格、品牌、評分 等)稀疏而導致的挑戰，這會 影響它們在預測較少互動項目時的表現。 為了提升整體效能，我們的研究專注於改進 CARCA 模型的項目嵌入層。 這項改進旨在更好地處理那些訓練不足的項目。我們使用了四個真實世界 的推薦系統數據集來評估我們的方法。研究結果顯示，我們的方法在預 測用戶可能感興趣的項目方面，比現有的先進模型更為優越。;Recommendation systems have achieved significant success in the field of deep learning, particularly for items with abundant interaction records. However, these systems often face challenges due to the sparsity of item attributes (such as price, brand, ratings, etc.), which hinders their performance when predicting interactions for less frequently engaged items. To address improve overall perforrmance, our research focuses on improving the item embedding layer of the CARCA model. This enhancement aims to better handle items that have not been adequately trained. We evaluated our approach using four real-world recommendation system datasets. The findings suggest that our method provides superior predictions of items that users may find interesting compared to the current state-of-the-art models.
