Supplementary material for Subtle Topic Models and Discovering Subtly Manifested Software Concerns Automatically
Mrinal Kanti Das, Suparna Bhattacharya, Ibm Research-India, Chiranjib Bhattacharyya, Kanchi Gopinath · 2007
We plot the average recall (over all datasets) of the topics to check how performance of the models change by increasing the degree of subtlety (DoS), see Figure 1. We see that STM consistently performs better than all models in this range of DoS. FTM which is the best among the existing models perform quite steadily until DoS is less than 0.5, after that we see a slight declination. The ability of MG-LDA to detect local topics makes it better than HDP and the difference is more for higher DoS, where it also narrows down the gap with FTM. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1