Abstract

Computational social studies using public social media data have become more and more popular because of the large amount of user-generated data available. The richness of social media data, coupled with noise and subjectivity, raise significant challenges for computationally studying social issues in a feasible and scalable manner. Machine learning problems are, as a result, often subjective or ambiguous when humans are involved. That is, humans solving the same problems might come to legitimate but completely different conclusions, based on their personal experiences and beliefs. When building supervised learning models, particularly when using crowdsourced training data, multiple annotations per data item are usually reduced to a single label representing ground truth. This inevitably hides a rich source of diversity and subjectivity of opinions about the labels.

Label distribution learning associates for each data item a probability distribution over the labels for that item, thus it can preserve diversities of opinions, beliefs, etc. that conventional learning hides or ignores. We propose a humans-in-the-loop learning framework to model and study large volumes of unlabeled subjective social media data with less human effort. We study various annotation tasks given to crowdsourced annotators and methods for aggregating their contributions in a manner that preserves subjectivity and disagreement. We introduce a strategy for learning label distributions with only five-to-ten labels per item by aggregating human-annotated labels over multiple, semantically related data items. We conduct experiments using our learning framework on data related to two subjective social issues (work and employment, and suicide prevention) that touch many people worldwide. Our methods can be applied to a broad variety of problems, particularly social problems. Our experimental results suggest that specific label aggregation methods can help provide reliable representative semantics at the population level.

Library of Congress Subject Headings

Social sciences--Computer simulation; Human-computer interaction; Social media; Data mining

Publication Date

6-2020

Document Type

Dissertation

Student Type

Graduate

Degree Name

Computing and Information Sciences (Ph.D.)

Department, Program, or Center

Information Sciences and Technologies (GCCIS)

Advisor

Christopher Homan

Advisor/Committee Member

Cecilia Ovesdotter Alm

Advisor/Committee Member

Rui Li

Campus

RIT – Main Campus

Plan Codes

COMPIS-PHD

Share

COinS