EAGER: Toward Interpretation of Pairwise Learning

为了展示机器智能，系统不仅要做出智能决策，而且要能够解释它是如何做出这些决策的。可解释的机器学习是一个新兴的研究领域，旨在开发新的或修改的机器学习技术，以产生更多可解释的模型。该项目将开发新的方法来解释被称为“成对学习模型”的一类学习模型的预测，该模型预测实例之间的关系，而不是单个实例的特定属性。例如，消费者可能想知道为什么系统推荐产品A与产品B相似。这项研究的结果可能有利于许多现实世界中的应用，涉及成对学习，如人脸识别，视觉跟踪，信息检索和生物信息学。所提出的方法的发展将有助于探索可解释的机器学习以及一般的机器学习。进行两个探索性的研究任务，以产生对个人的预测（本地解释）和整个模型的行为（全球解释）的解释。建议的本地解释方法适应Shapley值的概念来解释关于任意一对输入实例的预测决策。所提出的全局解释方法是一种贝叶斯非参数成对解释方法与弹性网，这将解释整个人口的特征重要性。这种全局可解释性可以促进对目标成对学习模型对特定输入维度的敏感性水平的理解。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的知识价值和更广泛的影响审查标准进行评估来支持。

To exhibit machine intelligence, it is critical for the system to not only make intelligent decisions, but also be able to explain how it arrives at those decisions. Explainable machine learning is an emerging research area aiming to develop new or modified machine learning techniques that will produce more explainable models. This project will develop novel methods to interpret the predictions of a class of learning models known as "pairwise learning models," which predicts relationships between instances rather than specific properties of an individual instance. For example, a consumer might want to know why the system recommends product A as similar to product B. The results of this research may benefit many real-world applications that are involved in pairwise learning, such as face recognition, visual tracking, information retrieval and bioinformatics. The development of the proposed approaches will contribute to the exploration of explainable machine learning as well as machine learning in general. Two exploratory research tasks are carried out to generate explanations on both individual predictions (local interpretations) and the entire model behaviors (global interpretation). The proposed local interpretation method adapts the concept of Shapley-value to explain prediction decisions about an arbitrary pair of input instances. The proposed global explanation method is a Bayesian non-parametric pairwise interpretation method with the elastic nets, which will explain feature importance across a population. Such global interpretability can facilitate the understanding of the sensitivity levels of a target pairwise learning model to specific input dimensions.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.