Multi-label Text Classification

Requirements

• Python 3.8
• All the modules in requirements.txt

Before we can use NLTK for tokenization some steps need to be completed. Open a new python session and run:

import nltk
nltk.download('punkt')

The datasets

• GoogleNews Word embeddings
• AG new corpus
• AmazonCat-13K

Additionally, the Wikipedia Stats.

Preprocessing the AmazonCat-13K dataset

Ensure that the project folder has the following directory structure:

• datasets/
  • AmazonCat-13K/
    • tst.json
    • trn.json
  • GoogleNews-vectors-negative.bin.gz
• 09 - Preprocess the AmazonCat-13k Dataset.py

The datasets folder contains the extracted archive files. Run task 09 - Preprocess the AmazonCat-13k Dataset. This should create the following files in the datasets/AmazonCat-13K folder:

• X.trn.raw.npy (an ndarray)
• Y.trn.raw.npz (a csc sparse matrix)
• X.trn.processed.npy (an ndarray)
• Y.trn.processed.npz (a csc sparse matrix)
• X.tst.npy (an ndarray)
• Y.tst.npz (a csc sparse matrix)

Training classifiers with the AmazonCat-13K dataset

Ensure that the project folder has the following directory structure:

• datasets/
  • AmazonCat-13K/
    • X.trn.processed.npy
    • Y.trn.processed.npz
    • X.tst.npy
    • Y.tst.npz
  • GoogleNews-vectors-negative.bin.gz
• results/
  • history/
  • predict/
  • weights/
• 10 - Training the AmazonCat-13k Dataset.py

Before running the training, the dataset needs to be preprocessed. To train the models, run the task 10 - Training the AmazonCat-13k Dataset.

Important labels

The top 10 most frequent labels are

Label id	Label	# of occurences
1471	books	355,211
7961	music	194,561
7892	movies & tv	128,026
9237	pop	120,090
7083	literature & fiction	97,803
7891	movies	88,967
4038	education & reference	76,277
10063	rock	75,035
12630	used & rental textbooks	71,667
8108	new	71,667

Labels that occur at most

Label id	Label	threshold	# of occurences
6554	john	50	50
4949	fountains	100	100
7393	marriage	1,000	996
84	accessories & supplies	10,000	9,976
9202	politics & social sciences	50,000	48,521
7083	literature & fiction	100,000	96,012