French Web Domain Classification

Authors : ABBES Siwar, ABED Dina, Karray Elyess

Introduction

The goal of this project is to solve a web domain classification problem. We have 8 different categories: business/finance, education/research, entertainment, health/medical, news/press, politics/government/law, sports and tech/science.

Text based methods and Graph based methods on web pages document for classification tasks are of many studies. In this project, we implemented different prepossessing strategies to deal with our data and we explored different methods of machine learning and deep learning to solve a web domain classification problem. We present all the methods deployed and the best scored method.

Dataset

The dataset of French domains was generated from a large crawl of the French web that was performed by the DaSciM team. We are given the following files:

1. edgelist.txt: a subgraph of the French web graph. It has 28, 002 vertices and 319, 498 weighted, directed edges. Nodes correspond to domain ids and there is an edge between two nodes if there is a hyperlink from at least one page of the source domain to at least one page of the target domain.
2. text directory: for each domain, a .txt file containing the text of all the pages of the domain. The text was extracted from the HTML source code of the pages.
3. train.csv: 2, 125 labeled domain ids. One domain id and category per row. The list of categories is shown in Table 1.
4. test.csv: 560 domain ids the category of which is to be predicted. One domain id per row.
5. graph baseline.csv: output of the provided graph baseline.

Preprocessing

To deal with our data (text data and graph data) and make them ready for the models, we did:

• Text feature extraction
• Text cleaning
• Lemmatization
• Word Embedding
• Document Embedding
• Graph feature extraction

Models Used

We first used different methods to get the encoding and add new features and then we used, in particular, three methods of classification: CNN, XGboost and CamemBERT.

Evaluation

The metric for this project is the multi-class logarithmic loss. It is defined as the negative loglikelihood of the true class labels given a probabilistic classifier’s predictions.

Results

The report "project_report.pdf" contains the results that we had for testing the different models mentioned in the previous section and more details about this project.

Code description

This project is composed mainly in three directories:

• models: where we can find the different models we used.
• utils: where we find util functions for data loading and preprocessing.

We made scripts to make our script more usable. You could run the following steps after preparing the data folder:

• Install deps.sh to install dependencies using this command: $ sh install deps.sh
• cross val.py in order to run cross-validation (change model in cross val.py) using this command: $python cross val.py
• Write submission.py to write submissions with the CNN model using this command: $ python write submission.py
• Write submission stacked.py in order to write submissions with the stacked models using this command: $ python write submission stacked.py