Measuring Time-Sensitive and Topic-Specific Influence in Social Networks with LSTM and Self-Attention.

Area of Science:

• Social Network Analysis
• Computational Social Science
• Information Science

Background:

• Influence measurement is crucial for applications like marketing and political campaigns.
• Existing methods struggle with dynamic social networks and topic-specific influence.
• Multi-typed interactions in social media complicate influence modeling.

Purpose of the Study:

• To propose a novel method for measuring time-sensitive and topic-specific influence.
• To address limitations of static models in dynamic, multi-topic environments.
• To develop a model that accounts for varying user influence over time and across subjects.

Main Methods:

• Developed the Time-sensitive and Topic-specific Influence Measurement (TTIM) method.
• Utilized a self-attention mechanism to model diverse interaction types.
• Employed matrix-adaptive long short-term memory to track influence dynamics.
• Integrated streaming text data with dynamic social network structures.

Main Results:

• TTIM successfully models time-sensitive and topic-specific influence.
• The model demonstrates superior performance compared to state-of-the-art methods.
• Experimental results on Twitter and Reddit datasets validate the approach.
• The method supports online learning with constant training time.

Conclusions:

• TTIM is the first method to measure time-sensitive and topic-specific influence.
• The model offers potential for visualizing influence dynamics and topic distributions.
• This approach enhances understanding of influence in complex, evolving social networks.