Aihemallinnus sekä muut ohjaamattomat koneoppimismenetelmät yhteiskuntatieteellisessä tutkimuksessa: kriittisiä havaintoja
Keywords:
aihemallinnus, ohjaamaton koneoppiminen, koneoppiminen, laskennallinen yhteiskuntatiede, puolueohjelma, massapuolueAbstract
Topic modelling is an unsupervised machine learning technique. Its potential applications for social science have increased during the recent years, both in Finland and internationally. Topic models, like many other unsupervised machine learning methods, requite input from the researchers in the form of parameters used with these models. Through a user study, this article demonstrates that the commonly used interpretability-focused approaches can lead to different outcomes. Also differences emerging from parameter selection are discovered in the outcomes of the analysis. Based on the experimental and empirical study, the article recommends that (i) the choice of parameters should be done using statistical measures instead of interpretability, and (ii) the results are further elaborated using social science literature or extended analysis. Furthermore, it is recommended that (iii) there should be more transparency related to the application of unsupervised methods, and (iv) researchers applying computational methods must follow the developing research on critical algorithm studies.
