Knowledge Intensive Learning: Combining Qualitative Constraints with Causal Independence for Parameter Learning in Probabilistic Models

Abstract

Medical problems are examples of knowledge-rich and data-poor domains. There is abounding knowledge about the target features ascribing to the centuries of medical research while the positive samples are rare due to the peculiarity of some diseases. Because the training data is so sparse, a learning algorithm that builds predictive models must be able to exploit the availability of such rich domain knowledge. However such medical knowledge is rarely employed when using machine learning for building predictive models. Such knowledge has been limited to identifying the "features" (i.e., attributes) that are useful in the target prediction. In this work, we provide a methodology of exploiting prior knowledge when learning probabilistic models.