Transformers compete on imbalanced product classification data

Context:

• The task chosen is to classify product data into its respective "browse node"

• Product data such as "Product name", "Product description", "Brand Name", "Brand Description" is given.

• As seen in E-commerce websites, The number of products existing in each category is highly varied, and this imbalance makes categorization by machine-learning challenging, and interesting.

• Hence, the performance of the amazing transformer models are explored on this imbalanced data.

details

• The dataset can can be found in this link. [link]

• link to resulting publication. [link]

Steps taken in this project:

1. The dataset is explore and pre-processed. [EDA and preprocessing]

2. The pre-processed dataset is tokenized using rust based batchwise tokenization and saved as pickle files. [tokenization]

3. Each transformer is trained using its respective tokensized dataset, the metrics are logged, and then visualized as shown below. [training] :
   
   

4. The top performer is retrained using the focal loss, and the improved is shown as below.

---

References

[1]  Surana S, "Amazon Product Browse Node Classification Data", Kaggle Datasets. [link]