this is for holding javascript data
Denes Csala added THe_other_weak_point_of__.md
over 8 years ago
Commit id: 93e15bd8918ab1b5f0562b50a677c9f3fb29b79f
deletions | additions
diff --git a/THe_other_weak_point_of__.md b/THe_other_weak_point_of__.md
new file mode 100644
index 0000000..2da5556
--- /dev/null
+++ b/THe_other_weak_point_of__.md
...
THe other weak point of the paper is the mathematical justification - or the lack thereof - of the AMC algorithm. While the M-GPU is a very elegant description, it is hard to keep track why does it work and the scale of the assumptions of not determined. For example, in equation 7 of MTD, the choice of the parameter beta is crucial: very large values will render the rest of the parameters insignificant.
The results of the paper are promising and the data sample presented to validate the results is impressive. However, one of the Achilles heels is that the aUthors do not include any asymptotic complexity derivations, albeit they claim that the algorithm runs in polynomial time. Although only qualitative, I particularly liked the assessment of the results by human judges. The advantage of the AMC over the previous best LTM was only marginal in the topic coherence metric - meaning that the LTM created just as coherent topics as did the AMC - with slightly different words at first look, in the second human test - word binning - the advantage was obvious. ALthough the weakness of the qualitative human test comes surfaces at this point: one must not forget that the topic label is arbitrary, only put by the human, the machine only sees a word distribution. Choosing a different topic label can actually greatly influence the word binning accuracy results - rendering this metric de facto useless.
