On Combining Knowledge-Engineered and Network-Extracted Features for Retrieval

The quality of case retrieval in case-based reasoning (CBR) systems depends on assigning appropriate case indices. Defining feature vocabularies for indexing is an important knowledge acquisition problem for CBR, often addressed by hand. The manual process may result in high-quality vocabularies, but at considerable effort and expense, and it may be difficult for non-symbolic input such as images. Recently, the ability of deep learning (DL) to identify important features has made it appealing for learning to assign case features. However, such methods may miss features apparent to knowledge engineers. This paper presents a case study on methods for combining benefits of both engineered and DL-generated features. It considers case-based classification of cases described by both symbolic features and images. It evaluates the power of both types of features individually, examines how quality of engineered feature information affects their combined benefit, and tests network methods to generate weights for their combination. Experimental results show that in the test domain under suitable circumstances, the combined approach can outperform either method individually.