Using support vector machine ensembles for target audience classification on Twitter.

Area of Science:

• Social Media Analytics
• Machine Learning
• Computational Social Science

Background:

• Social media content is vast and diverse, making customer identification challenging for businesses.
• Automated methods are needed to efficiently analyze social media data for marketing purposes.

Purpose of the Study:

• To investigate unsupervised and supervised learning for target audience classification on Twitter.
• To minimize annotation efforts in identifying potential customers.
• To enhance business advantage in the social media landscape.

Main Methods:

• Utilized Twitter Latent Dirichlet Allocation (LDA) for unsupervised topic discovery from follower content.
• Employed a Support Vector Machine (SVM) ensemble trained on diverse topic domains.
• Applied bootstrapping for over-sampling in training dataset construction, comparing it to random sampling.

Main Results:

• Achieved high accuracy in target audience classification using the proposed methods.
• Demonstrated that bootstrapping improves classifier performance in SVM ensembles compared to random sampling.
• Validated the effectiveness of the ensemble system in leveraging data diversity.

Conclusions:

• The developed ensemble system accurately differentiates prospective customers from the general audience on social media.
• Statistical inference methods like bootstrapping enhance classifier performance for audience segmentation.
• This approach provides a significant business advantage in crowded social media environments.