- 學位論文
用於對比學習推薦系統的可微分自動化資料擴增方法
Differentiable Automatic Data Augmentation for Contrastive Learning Recommendation
摘要
對比學習已被證明是一種對於訓練穩健的序列式推薦系統非常有效的策略，並且持續達到最先進的性能。在對比學習框架內，先前的研究主要集中如何分辨與產生有效的序列資料擴增方法，而本研究透過提出一個全面的資料擴增框架，引入一種新穎的方法。此方法是一種實例級、與資料本身相關，並可學習的擴增資料選擇器。我們方法的核心機制是能夠從各式各樣的資料擴增手段中，選擇適當的策略。這一機制顯著提升了模型學習過程，確保每次套用的資料擴增手法，是針對特定任務和資料點進行量身定制的，進而使推薦系統的有效性和準確性提高。透過大量實驗，我們驗證了所提出解決方法的有效性，展示了其對提升序列式推薦模型性能的能力。
並列摘要
Contrastive Learning (CL) has proven to be a highly effective strategy in training robust sequence recommendation models, consistently achieving state-of-the-art performance. While previous research has primarily focused on identifying effective methods for generating augmented data sequences within the CL framework, this work introduces a novel approach by proposing a comprehensive framework for finer-grained augmentation. This framework facilitates instance-level, data-dependent, and learnable selection of augmentation sequences. At the core of our approach is a mechanism that enables the selection of appropriate augmentation strategies from a diverse set of options. This mechanism significantly enhances the learning process by ensuring that each augmentation is tailored to the specific task and instance, thereby contributing to improvements in system efficiency and accuracy. Through an extensive array of experiments, we validate the effectiveness of our proposed solution, demonstrating its capability to elevate the performance of sequence recommendation models.