appendix.md

Appendix for the Journal Article - Experimental study of Rehearsal-based Incremental Classification of Document Streams

Note: The full article is currently under review.

Appendix A

Results for the Experiments on the Private Dataset for Most-Frequently Occuring Classes Scenario

This section contains detailed results for different numbers of iterations against instances from existing classes (IEC) values for the experiments performed on private dataset trained using most-frequently occurring classes first. The 3 subplots in the top rows, and the two left-most subplots in the bottom rows of each figure present the rolling average and average accuracies for old, new, and all classes for a particular training iteration. The right-most subplot in the bottom row for each figure shows the model performance on static test set.

Results for the experiments where the number of base classes = 2, and the number of new classes is 1.

Results for the experiments where the number of base classes = 2, and the number of new classes is 2.

Results for the experiments where the number of base classes = 3, and the number of new classes is 3.

Results for the experiments where the number of base classes = 4, and the number of new classes is 4.

Results for the experiments where the number of base classes = 5, and the number of new classes is 5.

Appendix B

Results for the Experiments on the Private Dataset for Least-Frequently Occuring Classes Scenario

This section contains detailed results for different number of iterations against instances from existing classes (IEC) values for the experiments performed on private dataset trained using least-frequently occurring classes first. The 3 subplots in the top rows, and the two left-most subplots in the bottom rows of each figure presents the rolling average and average accuracies for old, new, and all classes for a particular training iteration. The right-most subplot in the bottom row for each figure depicts the model performance on static test set.

Results for the experiments where the number of base classes = 2, and the number of new classes is 1.

Results for the experiments where the number of base classes = 2, and the number of new classes is 2.

Results for the experiments where the number of base classes = 3, and the number of new classes is 3.

Results for the experiments where the number of base classes = 4, and the number of new classes is 4.

Results for the experiments where the number of base classes = 5, and the number of new classes is 5.

Appendix C

Results for the Experiments on the Private Dataset for Random Classes Addition Scenario

This section contains detailed results for different number of iterations against instances from existing classes (IEC) values for the experiments performed on private dataset trained via random addition of new classes. The 3 subplots in the top rows, and the two left-most subplots in the bottom rows of each figure presents the rolling average and average accuracies for old, new, and all classes for a particular training iteration. The right-most subplot in the bottom row for each figure depicts the model performance on static test set.

Results for the experiments where the number of base classes = 2, and the number of new classes is 1.

Results for the experiments where the number of base classes = 2, and the number of new classes is 2.

Results for the experiments where the number of base classes = 3, and the number of new classes is 3.

Results for the experiments where the number of base classes = 4, and the number of new classes is 4.

Results for the experiments where the number of base classes = 5, and the number of new classes is 5.

Appendix D

Results for the Experiments on the RVL-CDIP Dataset for Random Classes Addition Scenario

This section contains detailed results for different number of iterations against instances from existing classes (IEC) values for the experiments performed on RVL-CDIP dataset trained via random addition of new classes. The 3 subplots in the top rows, and the two left-most subplots in the bottom rows of each figure presents the rolling average and average accuracies for old, new, and all classes for a particular training iteration. The right-most subplot in the bottom row for each figure depicts the model performance on static test set.

Results for the experiments where the number of base classes = 2, and the number of new classes is 1.

Results for the experiments where the number of base classes = 2, and the number of new classes is 2.

Results for the experiments where the number of base classes = 3, and the number of new classes is 3.

Results for the experiments where the number of base classes = 4, and the number of new classes is 4.

Results for the experiments where the number of base classes = 5, and the number of new classes is 5.