Shared Multi-View Data Representation for Multi-Domain Event Detection.

Area of Science:

• Computer Science
• Data Science
• Artificial Intelligence

Background:

• Internet platforms generate vast amounts of multi-domain data (online news, social media).
• Heterogeneous data views require effective alignment and representation learning for event discovery.
• Existing methods may not fully capture intrinsic structures across diverse data sources.

Purpose of the Study:

• To develop an event detection framework for discovering real-world events from multiple data domains.
• To address the challenge of heterogeneous multi-view data alignment and representation.
• To enhance the accuracy and effectiveness of real-world event discovery.

Main Methods:

• Utilized initial dictionaries of labeled data samples for multi-view data alignment.
• Devised a shared multi-view data representation (SMDR) model incorporating various constraints (shared representation, low-rank, local invariance, reconstruction error, dictionary independence).
• Employed class-wise residual models based on reconstruction residuals for event discovery.

Main Results:

• The proposed SMDR model effectively learns shared underlying structures across heterogeneous data views.
• Class-wise residual models accurately discover events based on learned data representations.
• Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed framework.

Conclusions:

• The developed event detection framework offers a robust solution for discovering real-world events from multi-domain online data.
• The SMDR model provides an effective approach for learning shared representations from heterogeneous data.
• The methodology shows significant promise for advancing event detection capabilities.