Hierarchical Categorization of Digikala Items

Overview

This project focuses on categorizing Digikala items hierarchically, from coarse to fine-grained categories across three levels. The classification model leverages ParsBERT, with a Multi-Layer Perceptron (MLP) trained on top of the CLS output token for final classification.

Dataset

The dataset consists of items from Digikala, an online marketplace. Each item is labeled with hierarchical categories to facilitate structured classification.

Dataset Overview

• Total Size: 529,322 items
• Unique Categories:
  • First Level: 11 categories
  • Second Level: 96 categories
  • Third Level: 682 categories

Methodology

The general approach is to add an MLP layer on top of the CLS token from the pre-trained ParsBERT model and fine-tune the model for each category level.

Multi-Level Classification Approach

Level 1 Classification

• Input: The name of the item.
• Output: The predicted first-level category.

Level 2 Classification

• Training Phase:
  • The model receives the concatenation of the item name and the actual first category.
  • It is trained to predict the second-level category.
• Testing Phase:
  • The Level-1 model is used to predict the first category.
  • The predicted first category is concatenated with the item name.
  • This concatenated input is passed to the Level-2 model to predict the second category.

Level 3 Classification

• Training Phase:
  • The model receives the concatenation of the item name, actual first category, and actual second category.
  • It is trained to predict the third-level category.
• Testing Phase:
  • The Level-1 model predicts the first category.
  • The predicted first category is concatenated with the item name and passed to the Level-2 model to predict the second category.
  • The item name, predicted first category, and predicted second category are concatenated and passed to the Level-3 model to predict the third category.

Results

The model achieves high accuracy in classifying Digikala items into their respective hierarchical categories. Detailed results and performance metrics are shown in the table below.

Level	Accuracy	F1 Score
Level 1	0.9843	0.9842
Level 2	0.9733	0.9732
Level 3	0.9576	0.9563

To evaluate the generalization performance of the model on another dataset, I crawled data from the Torob website for the first category and tested the model on this dataset. The results are shown below:

Dataset	Accuracy	F1 Score
Torob (Level 1)	0.9280	0.9424