使用個性化多互動偏好排名的多行為推薦系統

Abstract

多行為推薦的目標是利用用戶及物品的多交互關係例如購買和加入購物車來進行建模以解決推薦中常見的資料稀疏及冷啟動問題。雖然最近一些基於多行為的推薦演算法成功地利用不同種類的用戶及物品交互行為來提升推薦效果，但這些方法還存在一些限制。第一，大多數開創性的工作將單一行為當作目標行為且只根據目標行為來優化模型；然而，這需要重新訓練模型以預測其它行為，因此對於大規模資料集和許多實際應用來說效率很低。第二，雖然近期有些研究透過結合所有種類的行為一起進行模型優化以解決上述問題，但模型學習到的行為向量是所有用戶及物品都共用的，這樣的設定非常粗糙且不足以補抓用戶在不同行為下的偏好。除此之外，雖然這些最先進的多行為推薦演算法看似能對針對不同的行為對用戶推薦商品，但其它行為的預測並沒有被明確地評估在相關論文。我們透過使用個性化多互動偏好排名（PMiPR）來解決這些限制，它是應用於多行為推薦中有效及高效向量學習框架。具體來說，PMiPR 透過學習用戶及物品在每種行為下的特定行為向量將多行為信息整合至建模過程中。這不僅以更細粒度的方式對多行為信息建模，也讓我們能透過利用為指定的用戶及物品的行為向量來對不同的行為進行推薦。在四個公開的基準資料集上進行的綜合實驗證明了 PMiPR 在多行為推薦的有效性及效能。

The goal of multi-behavior recommendation is to leverage user-item interactions such as purchase and add-to-cart into the modeling process to address the commonly-faced data sparsity or cold start issues in recommendation. Although some recent multi-behavior-based recommendation algorithms successfully leverage different types of user-item interactions to improve recommendation performance, these methods still have limitations. First, most pioneering works treat a single behavior as the target behavior and optimize the model based on the target behavior only; this however necessitates re-training of the model to predict other behaviors and is thus inefficient for large-scale datasets and many real-world applications. Second, although recent studies address this issue by jointly optimizing the model based on all types of behaviors, the learned behavior embeddings are shared across all users and items, which is coarse-grained and insufficient to capture user preferences under different behaviors. Moreover, although such state-of-the-art multi-behavior recommendation algorithms seem able to recommend items for users w.r.t. different behaviors, they do not explicitly evaluate their methods in the reported experiments. We address these limitations with personal- ized multi-interaction preference ranking (PMiPR), an effective and efficient embedding learning framework for multi-behavior recommendation. Specifically, the proposed PMiPR incorporates multi-behavioral information into the modeling process by learning user-specific and item-specific behavior embeddings for each type of behavior. This not only models multi-behavioral information in a more fine-grained way but enables us to make recommendations w.r.t. different behaviors by leveraging the designated behavior embeddings for users and items. Comprehensive experiments on four public benchmark datasets demonstrate the effectiveness and efficiency of PMiPR for multi-behavior recommendation.