Incorporating additional information from our prior domain knowledge can be the key to solving difficult classification tasks, especially when the available training data is limited. The crucialstage of feature construction, often done manually, plays a significant role in allowing such information to be incorporated into a learner.We propose algorithms for automated feature construction where available domain knowledge,even though imperfect and approximate, can be utilized by the learning system. Robustness isachieved by incorporating this prior knowledge in a task-specific manner, guided by the actualtraining examples. These goals are realized with Explanation-Based Learning (EBL).The EBL paradigm provides the necessary bridge between domain knowledge and the trainingexamples, which allows us to design solutions that are conceptually well-formed and work for theright reason. The ideas of well-formed concepts and \working for the right reason" are our guidingprinciples for supervised learning.Using these underlying principles, we propose three algorithms for incorporating prior domainknowledge into discriminative learning with different levels of interaction between the feature construction process and the final classifier learning. The first approach involves automated construction of generative models for phantom examples, which can be used to enhance the training datafor subsequent classifier learning. Both the second and the third approaches involve the construction of semantic features. Each semantic feature encapsulates a well-formed concept which, ac-cording to the domain knowledge, corresponds to a conceptual difference between classes of objects.We illustrate and evaluate the proposed algorithms on the challenging problem of classifyingoffine handwritten Chinese characters, focusing on distinguishing difficult, mutually-similar pairsof characters. Empirical results show that our approaches can outperform the state-of-the-art algorithms.
PDF
download
878987797
028-85220240
