Research Article
A Novel Approach of Feature Space Reconstruction with Three-Way Decisions for Long-Tailed Text Classification
Table 4
The performance results of our proposed model and baseline methods.
| Method | Datatype | Accuracy | Macroprecision | Macrorecall | Macro-F1 score |
| TF-IDF | All data | 0.79 | 0.65 | 0.635 | 0.64 | Head class | 0.872 | 0.854 | 0.857 | 0.857 | Tail class | 0.635 | 0.597 | 0.581 | 0.585 |
| CNN | All data | 0.915 | 0.765 | 0.732 | 0.747 | Head class | 0.943 | 0.937 | 0.928 | 0.933 | Tail class | 0.747 | 0.731 | 0.692 | 0.709 |
| RNN | All data | 0.905 | 0.783 | 0.767 | 0.775 | Head class | 0.932 | 0.927 | 0.916 | 0.918 | Tail class | 0.782 | 0.759 | 0.728 | 0.742 |
| BI-LSTM | All data | 0.927 | 0.82 | 0.796 | 0.81 | Head class | 0.954 | 0.947 | 0.94 | 0.943 | Tail class | 0.821 | 0.778 | 0.747 | 0.752 |
| Our method | All data | 0.933 | 0.842 | 0.825 | 0.832 | Head class | 0.953 | 0.945 | 0.951 | 0.948 | Tail class | 0.841 | 0.813 | 0.797 | 0.801 |
|
|
Bold values means the best values in accuracy, macroprecision, macrorecall, and macro-F1 score.
|