Using Word N-Grams as Features in Arabic Text Classification - Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing - page 38

To train the CAL using the selected features from the training data, we used
three of the most commonly used CALs, KNN, naive Bayes (NB), and support
vector machine (SVM). For more information regarding FS methods, feature RSs,
and CALs, see [19]. We used the RapidMiner 4.0 [20] implementations of these
CALs to train and test the classification model. The experimental parameters are
summarized in Table 2.

3 Results and Discussion

Conducting experiments with the parameters in Table 2, we obtained the data in
Table 3.

Table 3 Experimental results

Classifier Grams Average Minimum Maximum

1 60.35 39.25 IG, LTC, 150 75 CHI, TFiDF, 200

2 50.38 25.88 CHI, LTC, 200 66.45 IG, Boolean, 200

NB

3 38.76 21.05 CHI, LTC, 200 51.32 DF, Boolean, 200

4 32.77 18.86 CHI, LTC, 200 42.98 GSS, Boolean, 200

1 49.14 40.13 DF, TFiDF, 100 58.77 CHI, LTC, 50

2 41.35 28.29 IG, Boolean, 50 51.54 CHI, LTC, 50

KNN

3 37.63 33.33 DF, TFiDF, 50 41.89 IG, LTC, 150

4 33.54 30.48 IG, Boolean, 50 36.4 GSS, LTC, 150

1 72.35 67.54 DF, TFiDF, 50 75.44 IG, LTC, 200

2 58.73 49.78 DF, Boolean, 50 65.13 IG, LTC, 200

SVM

CHI, Boolean, 50

3 41.81 35.53 47.37 IG, TFiDF, 200

4 35.07 31.58 CHI,Boolean,50 38.6 GSS, LTC, 100

The table lists the average CA for each gram number of each classifier, the
minimum and maximum CA for that classifier, and the combination of FS method,
RS, and number of terms that has produced the respective value. The data suggest
that, on average, for all gram numbers, the CA of SVM is greater than that of NB,
followed by that of KNN. In addition, the data suggest that greater CA is achieved
when single words are used as a feature, and that CA declines by 17%, on average,
when the gram number increases.
Notably, while SVM achieved greater CA, on average, and the best CA using
single words, NB exhibited greater CA for 2-grams, 3-grams, and 4-grams. The
data show that the best CA was achieved when the number of terms was 200, the
maximum number of terms used in our experiments. According to the data, NB
worked well with Boolean representation (three of the best results were achieved