Predicting the popularity of online content is highly valuable in many applications and has been studied for several years. However, existing models either work in population level---all messages are assumed to follow similar popularity dynamics, lacking flexibility to capture the intrinsic complexity of popularity dynamics, or work in individual level---the popularity dynamics of messages are independent of each other, failing to leverage other messages to improve prediction accuracy. In this paper, we propose a divide and conquer framework for popularity prediction. We first divide messages into groups, anticipating each group of messages follow similar popularity dynamics, and then, we train a group-specific model for the messages of each group. Experiments demonstrate that group-specific models improve the population-level models by about 30% and outperform state-of-the-art individual-level model.
Paper
Full text
Predicting the Popularity of Online Content with Group-specific Models
Semantic Scholar · Computer Science · 2017
Abstract
Predicting the popularity of online content is highly valuable in many applications and has been studied for several years. However, existing models either work in population level---all messages are assumed to follow similar popularity dynamics, lacking flexibility to capture the intrinsic complexity of popularity dynamics, or work in individual level---the popularity dynamics of messages are independent of each other, failing to leverage other messages to improve prediction accuracy. In this paper, we propose a divide and conquer framework for popularity prediction. We first divide messages into groups, anticipating each group of messages follow similar popularity dynamics, and then, we train a group-specific model for the messages of each group. Experiments demonstrate that group-specific models improve the population-level models by about 30% and outperform state-of-the-art individual-level model.