Multilabel text classification

1. Data file you can find in src/data folder. There is multiclass classification problem with 4 classes in total. Classes are balanced.

2. I trained 2 models: baseline (tf-idf + catboost) and advanced (bert). Training pipelines in src/train_tfidf_catboost.ipynb and src/train_bert.ipynb notebooks. Both models trained on same train data and evaluated on the same test data.

3. Confusion matrix catboost:

4. Confusion matrix bert:

5. Tf idf + catboost has 11 misclassified objects and bert 5 out of 2000 test samples. That was quite predictable.
6. I implemented API on flask with catboost model. To launch in use docker container:

docker build . -t bert:latest

docker run -it -p 8889:8889 bert:latest

6. Request API according to this example

curl --request POST 'http://localhost:8889/get-category' \
--data-raw '{
    "main_text": <some text>,
    "add_text": <some text>,
    "manufacturer": <some text>,
}'

7. Production code. Tf-idf model is nearly ready for production. Model itself can we enhanced a little bit by searching more optimal hyperparameters. Also it's better to implement API with an asynchronous API framework like Tornado. In this way, API will work slightly faster. You should deploy several model pods (better with autoscaling) to fit the desired number of requests. BERT model requires several additional steps: you should decide where to inference it: GPU or CPU. GPU is quite expensive. To inference on CPU it's better to use openvino before deployment. The rest steps is similar to tfidf model.