10.6: Conclusion
For creating a machine learning
model, no guaranteed method works. This is especially true for text
classification models. Using a variety of feature engineering techniques can
improve the odds of being able to create models with high predictive
performance. In the case of text classification, features are in thousands.
More the features, the more the model complexity. Complex models lead to higher
computational costs to train and use such models. Feature selection can help us
reduce the number of features and model complexity. We can also use feature
reduction to reduce the size of the feature matrix, without losing information
from features. This can reduce model complexity even further.
Many times, we find different
modeling techniques perform differently on the same test data. To harness the
power of different models, we can ensemble. The drawback of ensembling is that
it will lead to a huge number of base models. This results in higher
computational costs for training and using the model.
Ensemble feature selection can help
solve all of the above problems by doing feature selection for each base model
to identify fewer complex models for different combinations of features,
feature representation, and models. Finally, it identifies the best set of
models which give the highest predictive performance. Thereby reducing the
number of base models.