Machine-based mapping of innovation portfolios

Machine learning techniques show a great promise for improving innovation portfolio management. In this paper we experiment with different methods to classify innovation projects of a high-tech firm as either explorative or exploitative, and compare the results with a manual, theory-based mapping of these projects and with expert classification. We find that by combining a high-information extraction method with a decision tree or maximum entropy algorithm, higher levels of accuracy can be reached. Opportunities and limitations of different methods are discussed.