↓Skip to main content

Wikipedia Web Traffic Prediction

Table of Contents

Research Paper (Under Revision) | Co-authored with Prof. Sandhya Avasthi, ABES Engineering College, India | 2025

Overview

This project conducts an empirical benchmarking study across seven time-series forecasting algorithms evaluated against a massive 145,000-article Wikipedia web traffic dataset from Kaggle. The primary evaluation criterion was Root Mean Squared Error (RMSE) under varied temporal seasonalities.

Models Evaluated

  1. Moving Average (MA)
  2. Exponential Smoothing (ETS)
  3. Autoregressive Moving Average (ARMA)
  4. Autoregressive Integrated Moving Average (ARIMA)
  5. Auto-ARIMA
  6. Facebook Prophet
  7. Long Short-Term Memory Networks (LSTM)

Key Empirical Findings

Contrary to prevailing assumptions that recurrent neural networks (LSTMs) consistently dominate tabular sequence problems, Moving Average models demonstrated the strongest generalization and lowest overall RMSE for this high-variance web article distribution. This paper underscores the crucial importance of empirical model selection, showing that simpler, interpretable statistical baselines can often outperform computationally expensive deep learning models on large-scale noisy time series.

International Collaboration

This research represents a successful academic cross-border collaboration between Batangas State University (Philippines) and ABES Engineering College (India). The full paper has been submitted to a peer-reviewed journal and is currently undergoing revision.

Tech Stack: R · Time-Series Analysis · ARIMA · Prophet · LSTM · Kaggle GPU