Bank complaint text mining
Topic modelling and classifiers over 144,000 consumer bank complaints, plus a risk rating built from 692 hand-labelled cases.
Team member · SMU IS450 Text Mining
Result
TF-IDF classifiers reached a macro F1 of 0.85. On the small hand-labelled set, logistic regression (macro F1 0.80) beat DistilBERT. That is a small-data lesson I keep coming back to.
Built with
- Python
- scikit-learn
- LDA
- DistilBERT