CS412 Machine Learning Project: Turkish Instagram Influencers

Overview of the Repository

This repository contains the following key components for the machine learning project:

• CS412_Project_Report.pdf: A comprehensive report detailing the project's objectives, data description, methodology, results, and future work.
• classification_task.ipynb: A Jupyter Notebook implementing the classification task to categorize user accounts into different domains.
• regression_task.ipynb: A Jupyter Notebook implementing the regression task to predict the number of likes for user-generated content.

Methodology

The project tackles two main machine learning problems:

1. Regression Task: Predicting the number of likes for user-generated content using LightGBM.

   • Key features include historical user engagement metrics, post characteristics, and logarithmic transformation of the target variable.
   • The model demonstrated effective performance with a low Log Mean Squared Error.

2. Classification Task: Categorizing user accounts into domains using a Voting Classifier.

   • Preprocessing techniques include text cleaning, TF-IDF vectorization, and handling class imbalance with SMOTE.
   • The ensemble model, combining Logistic Regression, SVC, and XGBoost, achieved robust classification performance across various categories.

Results

Regression Task

• Log Mean Squared Error: 0.078
• Prediction range: 0 to 1,327,989 likes
• Number of zero predictions: 534 out of total samples

Classification Results

Category	Precision	Recall	F1-Score
Art	0.23	0.08	0.12
Entertainment	0.45	0.40	0.42
Fashion	0.53	0.72	0.61
Food	0.86	0.87	0.87
Gaming	0.00	0.00	0.00
Health and Lifestyle	0.62	0.79	0.70
Mom and Children	0.71	0.40	0.51
Sports	0.80	0.70	0.74
Tech	0.71	0.75	0.73
Travel	0.67	0.64	0.66

Note: The model performs well for categories like Food, Sports, and Tech, while struggling with smaller or more niche categories like Gaming and Art.

Team Contributions

• Efe Güçlü
• Irmak Sözen
• Bora Sacır
• Murat Ozan Aybak
• Janset Tunca