To Do:

[amritbhanu]

For Research

• Does stability help classification? Tuned and untuned? Multigoal fscore and jaccard score
• IMP: Ldaga paper reproduction (https://dibt.unimol.it/reports/LDA-GA/)
• And topic matching with lda and with jist brute and force 7,10 words

For LN

• Tune svm and lda, just for fscore (kernel, C, degree, learning rate)
• Lda+word vector which will have topic word features (Part of my research as well)
• Mutate Tags by x% to see what error it throws. (Heat Map with tags and topics for different datasets).

Conclusion:

• Feature engineering, LDA features are much better for text mining rather than tf, tfidf. (Can be a paper)

[amritbhanu]

• tfidf weight + kmeans clustering (Raymond Mooney (Learning Scripts for Text Understanding with Recurrent Neural Networks) seminar title)
• lda features + kmeans

LN:

• Std error types
• inter document similarity
• long documents, ambiguous, which makes it confusing
• cosine, with some threshold
• tar3 learner contrast
• http://unbox.org/stuff/tags/tar/3.0/c/

[amritbhanu]

PAPERS WHICH CAN BE MADE OUT OF IT:

• LDA + Classification
• LDA features compared with tf , tfidf (Compared with other text mining results) and it performs better with DE
• LDADE results compared to LDAGA

[timm]

"LDA features compared with tf , tfidf (Compared with other text mining results) and it performs better with DE" is probably a small section in external validity

lets see what icse says but i think you have one paper showing that stability is useful for supervised and unsupervised learning

i long to want your stats tests on f2, f1, jacard before and after tuning.

[amritbhanu]

• magic parameters tune. New set of examples they want surely classified and not surely classified.
  • waiting for results