Supervised and Unsupervised Feature Selection for Inferring Social Nature of Telephone Conversations from Their...

Area of Science:

• Computational Social Science
• Natural Language Processing
• Human-Computer Interaction

Background:

• Understanding the nature of telephone conversations has broad applications, including context-aware smartphone interfaces and social science research.
• Existing methods for analyzing conversations often require extensive feature engineering or supervised approaches.

Purpose of the Study:

• To investigate the effectiveness of content-based features in classifying business-oriented telephone calls from other types of calls.
• To evaluate the utility of unsupervised learning methods for inferring the social nature of conversations.

Main Methods:

• Utilized a unique corpus of everyday telephone conversations collected over one year from eight residences.
• Employed feature selection experiments to identify key discriminators between call types.
• Applied unsupervised methods, specifically latent Dirichlet allocation (LDA), and compared their performance to supervised methods.

Main Results:

• A small set of content-based features robustly classifies the majority of telephone calls.
• Unsupervised learning features, particularly those from LDA, achieve performance comparable to supervised methods.
• Learned unsupervised clusters show potential for detailed inference of conversational social characteristics.

Conclusions:

• Content analysis of telephone calls is a viable method for classification and social inference.
• Unsupervised learning provides an efficient and effective alternative to supervised methods for analyzing conversational data.
• This research opens avenues for advanced social science research and context-aware technology development.