2017 Volume 68 Issue 3 Pages 161-170
In the development of the new product and the management of technologies, it is important to make decisions based on the information about technological trends, such as patents. The relationship between the application range and field of technology has become complicated due to the use of increasingly sophisticated technologies in recent years. Along with this, knowledge sharing between decision-makers and engineers has become increasingly difficult. Under these circumstances, there is a need for an efficient tool capable of supporting knowledge sharing among engineers and managers whose expertise are in different areas. The purpose of this study is the knowledge extraction of patent data through constructing a text-mining patent map using the random forest methodology. By extracting the knowledge from the results learned using the technological information obtained through random forest as the supervised signal, we have constructed patent maps that take the relationships between technical fields into consideration. The method we propose creates a patent map by combining two-dimensional kernel density estimation and multi-dimensional scaling based on a similarity matrix calculated applying the internal components of the random forest. We also deal with calculation results of non-negative matrix factorization as a random forest input, thereby avoiding the vulnerability of the random forest noise variables. Non-negative matrix factorization is useful for interpreting the important variables extracted in the random forest. In the experiment, we confirmed the behavior of the method proposed using Japanese patent data.
