Speaker : YI-CHENG HUNG Curriculum Learning for Heterogeneous Star Network Embedding via Deep Reinforcement Learning Source : WSDM 2018 Advisor : JIA-LING KOH Speaker : YI-CHENG HUNG Date:2018/04/24
Outline Introduction Method Experiment Conclusion
Outline Introduction Method Experiment Conclusion
Introduction-motivation 訓練模型 預測類型: 奇幻類型 NN的INPUT是embedding vector，要如何訓練一個足夠代表input的embedding vector? Embedding vector
Introduction-goal 如何得知最好的修課順序? Goal:在考試時拿到最好的成績 修課順序
Outline Introduction Method Experiment Conclusion
Method-framework 假設已經一經上過(A,A,B,B,C)的課程了 如何決定下一次的課程?
Method-Planning Module
Q-Learn η是學習率，決定此次誤差有多少被學習 Γ是折扣率，決定未來策略的比重 現實 估計
Method-framework
Method-Learning Module
Method-Integrating
Outline Introduction Method Experiment Conclusion
Experiment-dataset 多標籤類型會進行排除
Quantitative results in the unsupervised setting
Quantitative results in the semi-supervised setting
Performance w.r.t. training steps on the DBLP dataset.
Performance v.s. the running time.
Performance w.r.t. #simulations on the IMDB dataset.
Outline Introduction Method Experiment Conclusion
Conclusion Proposed a deep reinforcement learning based approach Integrates the learning and planning strategies Experimental results proved the effectiveness and efficiency of our approach.