LDA topics as feature selector

[amritbhanu]

Experiment rig:

• The initial run was with default parameter of k=10. Documents had 10 features. The score was null.
• neg/pos ratio was very high. Unbalanced dataset. Eg:
  • SE0: 'no': 6008, 'yes': 309 - we got Fscore of about 0.5
  • SE1 'no': 47201, 'yes': 1441 - we got Fscore of about 0.8
  • SE3 no': 83583, 'yes': 654 - 0 Fscore
  • SE6 'no': 15865, 'yes': 439 - 0 Fscore
  • SE8 'no': 58076, 'yes': 195 - 0 Fscore
• Still tuning experiments are running.

  Change in experiment

• Will try smote.
• And also i can change this default parameter k in the steps of 20, 40, 80.

[amritbhanu]

• [x] smallest no of terms, and stability which we will need and high f scores. (multi objective)
  • (maximized (fscore+Raw score))
  • Features (documents distributed over topics.) and smoting expecially for SE datasets.
  • RESULTS VERY BAD DUE TO very less column features. Values in 0.1
• [x] topics distributed over words. Results
  • SE, not done tf-idf yet. No of cols are very high. (maximized (fscore+Raw score))
  • other can be just using hashing trick.
  • choose the top most topic and then use weights of words.
• [ ] tuned k=20, with 7 words each. untuned k=40-100, with 10 words.
  • Check for the set overlap and simple matching and unmatching.
• [x] change the default parameter k in the steps of 20, 40, 80.
• [ ] Tune learners with tuned LDA. Objectives??
• [x] Randomness Results