Handling of Imbalanced Data in Text Classification: Category-Based Term Weights - Natural Language Processing and Text Mining - page 178

we are very concernedabout the performance of every category and especiallythe
minor ones, we report theoverall performance in macro-averaged scores, to avoid
the bias formajor categories in imbalanced data associatedwithmicro-averaged
scores.
10.6 Experimental Results and Discussion
10.6.1 Overall Performance
Figure 10.3 shows theoverall performance ofatotalof16weightingschemes tested
over MCV1 and Reuters-21578 . Theyare reportedintermsofmacro-averaged F 1
values.
Fig. 10.3. The macro-averaged F 1 values of16weightingschemes tested overMCV1
and Reuters-21578
Our first observationis that all TFFVweightingschemes outperform classic
ones, i.e., TFIDF, 'ltc,' and normalized'ltc' schemes. The TFIDF's performance
based onReuters-21578 is in line with theliterature[41, 46]. This hasshown the
overall effectiveness ofTFFV-based schemes. In general, the performance patternsof
all 16 schemes on MCV1 and Reuters-21578 matchvery well. For example, CBTW 1
and CBTW 3 always take thelead among CBTWs, normalized'ltc' always performs