This study will test the feasibility of alternative machine learning strategies for training an expert decision support computer program to automatically code social interaction processes. Collecting and analyzing interaction process data is difficult, time consuming, and expensive. Reliability and reproducibility are difficult to establish and maintain. Phase I will test several alternative machine learning strategies on health care practitioner - patient interaction data. We will assess the percent of events which can be automatically coded, the number of cases required for training, and the utility of alternative learning strategies. In Phase II the most promising strategies will be incorporated into a computer program for interacting coding. Automated interaction coding with acceptable rates of accuracy could drastically reduce the cost and time required for interaction analysis, improve reliability, enhance scientific accountability, and permit new applications in research and training. For research it could permit the analysis of far larger quantities of interaction data, permitting interaction studies to examine a broader range of variables and employ more sophisticated analyses. Coupled with advances in speech recognition it could permit real-time interpretation and advice for teachers, physicians, police, and other professions. In a service-oriented economy with increasing professionalization of many service providers, automated coding of interaction processes could provide a broad range of intelligent computer assistants to improve interaction skills.
