提升核机器解释性能的理论与方法研究

How to automatically improve interpretation and generalization performance of kernel Machine is an essential issue for domain experts’ trust in the kernel Machine. However, the trade-off problem between the generalization performance and the interpretability of the model is always a huge challenge. For the trade-off problem, an important work is to bring professional guidance into machine learning by constraining learning model. In order to do that, the following three works will be finished. In the first work, on account of characteristic of the incomplete fuzzy relationships between input attributes and output attribute in the practical engineering, the project will propose the representing method of the relationships in first-order predicate logic and functional forms, and an interpretation-oriented attribute reduction and amendment method will be proposed. In the second work, for evaluating the interpretability of the learning model, the evaluation criteria will be proposed. The project will research the affection of various constraint mechanisms in the conventional regularization learning framework on the trade-off problem. For the trade-off problem, we will research and propose an interpretation-oriented regularization learning framework and a theory model based on global soft constraint mechanism. In the final, based on the previous research results, the project will propose an interpretation-orient multi-task or single-task learning algorithm based on the global soft constraint for improving the interpretability of multi-task or single-task learning model. The algorithms will be applied in many practical engineering. In conclusion, the research results of the project will indicate the interaction among interpreted risk, empirical risk and expected risk and propose the solution of the trade-off problem for domain experts’ trust. It also will provide the basic theory support for the solution and extend the functional scope and applicable fields of the kernel method.

自动提升核机器解释性能是增强领域专家对核机器学习信任的一个重要策略。其根本问题是学习模型解释性能与泛化性能之间的均衡问题。考虑到在学习过程中引入专业指导并进行合理约束，是提升核机器解释性能的重要一环。故本课题拟从以下三个方面进行探索。第一，实际工程中先验知识通常为不完备的模糊属性关系，故首先拟引入一阶谓词定义多种基本关系，然后利用马尔科夫逻辑网不确定性推理对其进行修正，最后引入先验函数向量和Tnorm函数给出关系的数学表达式；第二，针对该均衡问题，拟提出基于全局弱约束机制的正则化框架来约束学习过程，并基于统计学习理论研究两种性能的制衡机理，给出该方案的理论依据；第三，实现提升单任务和多任务核机器解释性能的方法，并将理论与算法成果用于实际应用中。所获结果将有望解决解释性能与泛化性能均衡问题，增强领域专家对核机器学习算法的信任。在应用层面将有望扩展现有核机器学习算法的功能范围与可应用性领域。

自动提升核机器解释性能是增强领域专家对核机器学习信任的一个重要策略。其根本问题是学习模型解释性能与泛化性能之间的均衡问题。在学习过程中引入专业指导并进行合理约束，是提升核机器解释性能的重要一环。本课题从以下三个方面进行了探索研究。第一，实际工程中先验知识通常为不完备的模糊属性关系，故首先拟引入一阶谓词定义多种基本关系，然后利用马尔科夫逻辑网不确定性推理对其进行修正，最后引入先验函数向量和Tnorm函数给出关系的数学表达式；第二，针对该均衡问题，提出了基于全局弱约束机制的正则化框架来约束学习过程，基于统计学习理论研究两种性能的制衡机理，给出该方案的理论依据，；第三，实现了提升单任务核机器解释性能的方法，并将理论与算法成果用于实际应用中。提出了可解释的最小二乘支持向量机、面向卷积神经网络判别部分黑箱问题的自解释方法和面向主动噪声控制的误差建模方法，研究了训练误差概率分布对泛化性能的影响和最优种群一致逼近的总体分布。所获结果将有望解决解释性能与泛化性能均衡问题，并为机器学习的可解释问题提供新的解决思路。