NTU » Semantic Search Engine

I am opening up this issue to track the progress of the TFKL paper.

As discussed, we need to

• redo the 2-class problems for TFRF, using 2 classifiers

• replot the curve for TF.KL using absolute value

• rewrite the intro, shifting the focus from sentiment analysis to supervised term weighting approaches, which is actually moving towards a feature-selection problem. You should also include some unsupervised term weighting approach like "A compartative study on feature selection in Text Categorization" by yiming yang and jan o pedersen.

• read and include some more related papers like:
  1. "an improved term weighting scheme for vector space model" by yue-heng sun pi-lian he, zhi-gang chen
  1. seach for "term absence" and text classification in google scholar

• emphasize the key advantages (selling point) of TF.KL
  1. it considers both term presence and term absence from the class distribution, and
  1. in the case of 2-class problems, only 1 classifier need to be trained, unlike tf.rf, which requires you to train C classifiers for C classes even when C=2.

• do more experiments on C>2

• try out TFKL on the (C choose 2) summation formula you showed me for C>2

• in the literature, compare with other supervised term weighting approach