Part 3 SPAM Naive Bayes Precision ham 98.79518072289156 Precision spam 95.64032697547684 Recall ham 98.4 Recall spam 96.69421487603306 F1 score HAM 0.9859719438877755 F1 score SPAM 0.9616438356164384 SVM_Light Precision ham 98.99497487437185 Precision spam 95.92391304347827 Recall ham 98.5 Recall spam 97.2451790633609 F1 score HAM 0.9874686716791979 F1 score SPAM 0.9658002735978113 MegaM Light Precision ham 98.94625922023182 Precision spam 96.08695652173913 Recall ham 93.89999999999999 Recall spam 60.88154269972452 F1 score HAM 0.9635710620831195 F1 score SPAM 0.7453625632377741

SENTIMENT Naive Bayes Precision pos 87.06896551724138 Precision neg 82.08955223880598 Recall pos 80.80000000000001 Recall neg 88.0 F1 score POS 0.8381742738589211 F1 score NEG 0.8494208494208494 SVM_Light Precision pos 87.87528868360278 Precision neg 87.10407239819004 Recall pos 86.97142857142856 Recall neg 88.0 F1 score POS 0.87421022400919 F1 score NEG 0.8754974417282548 MegaM Light Precision pos 75.34562211981567 Precision neg 74.94331065759637 Recall pos 74.74285714285715 Recall neg 75.54285714285714 F1 score POS 0.7504302925989673 F1 score NEG 0.7524188958451906

I chose a 10% of the spam training data for the purpose of testing. The following are the scores obtained: Naive Bayes Precision ham 97.61194029850746 Precision spam 94.6927374301676 Recall ham 98.1 Recall spam 93.38842975206612 F1 score HAM 0.9785536159600997 F1 score SPAM 0.9403606102635229 SVM_Light Precision ham 94.4980694980695 Precision spam 93.57798165137615 Recall ham 97.89999999999999 Recall spam 84.29752066115702 F1 score HAM 0.9616895874263262 F1 score SPAM 0.8869565217391305 MegaM Light Precision ham 98.34196891191709 Precision spam 95.6896551724138 Recall ham 94.89999999999999 Recall spam 61.15702479338842 F1 score HAM 0.9659033078880407 F1 score SPAM 0.746218487394958

There seems to be a slight decrease in the F-score of Naive Bayes classifier and the SVM_Light classifier. But not very significant. The reason for decrease could be that some of the prominent features from the whole training set are not present. However I was expecting a significant decrease in the F-score , but it could be because the training set chosen resembled the dev set very closely. On the contrary MegaM MaxEntropy classifier seems to have slightly better F-score, this is again strange to me. But the only possible explanation I can think of is that the training set possibly captured the features present in the dev set closely.