Hujjatullah Fazlurrahman, Bima Setyo Nugroho, Restu Eri Adinata, Novita Aprilia, Devangga Putra Adhitya Pratama, Thusy Tiara Saraswati, Reza Ayu Palupi, Noris Bagus Mulyo, A. Amirusholihin
This study evaluates the effectiveness of tree-based machine learning models for classifying dominant traffic sources of Instagram content using 29,999 posts characterized by engagement, exposure, and structural content metrics. Decision Tree, Random Forest, and Gradient Boosting algorithms were evaluated against random and majority baselines. Hyperparameter tuning and ablation experiments were conducted to examine model robustness and feature contribution. All models achieved accuracy levels near the theoretical random probability (≈ 0.167), with only marginal improvement. Ablation results indicate that neither engagement metrics nor structural content attributes substantially improve predictive performance. The findings suggest that traffic source dominance is structurally unpredictable when relying solely on standard Instagram analytics variables and is likely influenced by latent platform mechanisms. This study emphasizes the importance of benchmark baselines, dataset transparency, and cautious interpretation in social media predictive modeling. © 2026 IEEE.
Surabaya State University, Faculty of Economics and Business, Departemen of Digital Business, Surabaya, Indonesia; Graduate School of Management, Saint Petersburg State University, Departmen of Management, Saint Petersburg, Russian Federation; Surabaya State University, Faculty of Economics and Business, Departemen of Accounting Education, Surabaya, Indonesia; Surabaya State University, Faculty of Economics and Business, Departemen of Economics Education, Surabaya, Indonesia; Surabaya State University, Departemen of Office Administration Education, Surabaya, Indonesia; Surabaya State University, Digital Agribusiness, Faculty of Food Security, Surabaya, Indonesia
