Title: Identification of innovation diffusion trends with FDA clustering
Authors: Albina Latifi - Justus Liebig University Giessen (Germany) [presenting]
Peter Winker - University of Giessen (Germany)
David Lenz - Justus-Liebig University Giessen (Germany)
Abstract: The Diffusion of Innovation Theory often describes innovation diffusion as a hump-shaped curve. Light is shed on this theory by using a data-driven approach based on news articles from a technology-related newspaper for the period 1996 - 2021. In a first step, computational methods from natural language processing such as topic modelling were used to identify latent topics in the text corpus and to obtain associated time series of topic weights. In a second step, methods from the field of functional data analysis (FDA) were applied to categorize these time series in clusters. For this purpose, the $k$-means method, which is often used in the literature for related tasks was compared with an implementation of the global search heuristic Threshold Accepting (TA) for clustering. Preliminary results indicate that TA provides better and more robust results than k-means. Different clusters of innovation diffusion trends were identified, suggesting that empirically there are various shapes of diffusion which do not all resemble the classical diffusion curve to the same extent. Moreover, this approach could uncover different stages of innovation diffusion. Based on these results, success predictions for individual innovations might be derived.