LDA Ensembles for Interactive Exploration and Categorization of Behaviors

Area of Science:

• Computer Science
• Data Analysis
• Behavioral Science

Background:

• Analyzing large datasets of actor behaviors is crucial for understanding complex systems.
• Identifying typical and anomalous behaviors aids in system optimization and security.
• Traditional topic modeling, effective for text, faces challenges with non-textual action data.

Purpose of the Study:

• To develop an effective method for analyzing large collections of actor behaviors.
• To identify typical and anomalous behaviors within these collections.
• To adapt topic modeling for non-textual behavioral data.

Main Methods:

• Utilized Latent Dirichlet Allocation (LDA) Ensembles for behavior categorization.
• Developed visual and interactive techniques for topic construction and evaluation.
• Applied criteria such as distinctiveness and coverage for topic quality assessment.

Main Results:

• Demonstrated the effectiveness of LDA Ensembles in representing typical behavior categories.
• Successfully identified anomalous behaviors in case studies.
• Validated the approach through expert evaluation in a security management system analysis.

Conclusions:

• The proposed approach enhances the analysis of non-textual behavioral data.
• LDA Ensembles combined with interactive visualization offer a powerful tool for behavior analysis.
• The method is effective for identifying patterns and anomalies in diverse behavioral datasets.