Label | ε≤ | Time (mean) | Time (sd) | F 1 score (mean) | F 1 (sd) |
---|
flowEMMi | 1.0 | 528 | 53 | 0.56 | 0.18 |
flowEMMi | 0.01 | 1 080 | 214 | 0.59 | 0.17 |
flowEMMi | 10−5 | 1 445 | 182 | 0.56 | 0.17 |
flowMerge | 1.0 | 8 391 | 3 239 | 0.54 | 0.24 |
flowMerge | 0.01 | 8 951 | 3 597 | 0.51 | 0.17 |
flowMerge | 10−5 | 56 652 | 53 379 | 0.54 | 0.17 |
- Times and F 1 scores (and their standard deviation (sd)) are aggregated over four experiments and 5 expert user gatings, each. Note that the default flowMerge stopping criterion of 10−5 yields running times in excess of 1 day. flowEMMi consistently yields better F 1 measures with an average improvement of 4% to 16% over flowMerge, with much better running times, easily yielding speed improvements of ×8 – ×15 or better. For both algorithms, having a more stringent EM stopping criterion tends to increase the F 1 score, however especially for flowMerge at prohibitive running time costs