Why cannot long-term cascade be predicted? Exploring temporal dynamics in information diffusion processes.

Area of Science:

• Computational Social Science
• Network Science
• Machine Learning

Background:

• Information cascades are critical for applications like advertising and emergency management.
• Accurate long-term prediction of information cascade scale remains a significant challenge.
• Understanding diffusion dynamics is key to controlling information spread.

Purpose of the Study:

• To investigate the predictability of information cascade scales using machine learning on Weibo data.
• To compare the predictive power of various features (user network, tweet content, early dynamics) for short-term versus long-term prediction.
• To identify reasons for the decline in predictive power over longer time scales.

Main Methods:

• Extraction of diverse features from Weibo data, including user network, retweeting network, tweet content, and early diffusion dynamics.
• Application of conventional machine learning algorithms to predict information cascade scales.
• Comparative analysis of feature importance and predictive power for short-term and long-term tasks.

Main Results:

• Early diffusion dynamics were found to be the most predictive features for short-term cascade scale prediction.
• These early dynamics significantly lost predictive power in long-term prediction tasks.
• Feature temporal drift and the bursty nature of information diffusion were identified as key reasons for this predictive power loss.

Conclusions:

• The effectiveness of predictive features for information cascades diminishes over longer time scales.
• Bursty diffusion patterns and temporal drift in features challenge long-term prediction accuracy.
• Findings enhance understanding of information diffusion and inform strategies for controlling online information spread.