給定無法取得之特徵並學習預測

Abstract

從具有缺失值的數據中學習已成為現實應用程序中普遍面臨的挑戰。這項工 作著重於一個特定的場景，即在預測階段無法使用訓練特徵維的子集。在這種情況 下，大多數現有方法都存在指定特徵尺寸的空白，因此無法提供高質量的結果。這 項工作提出了一種新穎的基於神經的學習框架，以利用訓練中獲得的知識來減輕 預測過程中某些特徵缺失的影響。我們的解決方案結合了兩種知識轉移策略，使模 型可以從不斷減少的功能以及經過全面信息培訓的老師網絡中學習。實驗結果表 明我們的權重減輕算法的有效性以及我們的師生學習框架的整體優越性。

Learning from data with missing values has become a commonly faced challenge in real-world applications. This work emphasizes on a specific scenario that a subset of training feature dimensions becomes unavailable during the prediction stage. In this certain case, most of the existing approaches suffer from the vacancy of designated feature dimensions, thus not capable of providing quality results. This work proposes a novel neural-based learning framework to leverage the knowledge obtained during training to alleviate the effect from missing of certain features during prediction. Our solutions incorporate two knowledge transferring strategies allowing the model to learn from diminishing features as well as from a teacher network trained with full information. Experiment results show promising outcomes comparing with the state-of-the-art imputation-based solutions and the effectiveness of our weight diminishing algorithm and the whole superiority of our teacher-student learning framework, compared to state-of-the-art methods tackling missing data.