Publication Date

Spring 2016

Degree Type

Thesis

Degree Name

Master of Arts (MA)

Department

Linguistics and Language Development

Advisor

Hahn Koo

Keywords

Latent Dirichlet Allocation, Naive Bayes, Prediction, Reddit, Support Vector Machine, TF-IDF

Subject Areas

Linguistics; Computer science

Abstract

Popularity prediction is a useful technique for marketers to anticipate the success of marketing campaigns, to build recommendation systems that suggest new products to consumers, and to develop targeted advertising. Researchers likewise use popularity prediction to measure how popularity changes within a community or within a given timespan. In this paper, I explore ways to predict popularity of posts in reddit.com, which is a blend of news aggregator and community forum. I frame popularity prediction as a text classification problem and attempt to solve it by first identifying topics in the text and then classifying whether the topics identified are more characteristic of popular or unpopular texts. This classifier is then used to label unseen texts as popular or not dependent on the topics found in these new posts. I explore the use of Latent Dirichlet Allocation and term frequency-inverse document frequency for topic identification and naïve Bayes classifiers and support vector machines for classification. The relation between topics and popularity is dynamic -- topics in Reddit communities can wax and wane in popularity. Despite the inherent variability, the methods explored in the paper are effective, showing prediction accuracy between 60% and 75%. The study contributes to the field in various ways. For example, it provides novel data for research and development, not only for text classification but also for the study of relation between topics and popularity in general. The study also helps us better understand different topic identification and classification methods by illustrating their effectiveness on real-life data from a fast-changing and multi-purpose website

Download

Share

COinS