Retweetability Prediction

This work was submitted in partial fulfillment of Columbia University Master's Program

Original Thesis Title : The Use of Domain-Specific Sentiment Analysis on Predicting Information Diffusion in Online Social Networks

Short Summary

In this project, I built machine learning models (logistic regression, k nearest neighbor, random forests, and XGBoost) including domain-specific sentiment analysis that classify whether a message is retweetable or not --- with the highest AUC of 0.84.

Previous Work

This project was inspired by previous research projects I was involved in. If interested, here are some research findings that I presented at national social science conferences:

1. March 2018, the Society for Personality and Social Psychology Annual Meeting
2. May 2019, the Association for Psychological Science Annual Convention

Motivation

As online social media networks have become a major platform for sharing one’s opinions, there is a growing need for building an accurate predictive model for information diffusion. Taking insights from social science research (See the previous work for more details), I hoped to build a prediction model that incorporates a number of social factors associated with sharing political content in online social networks.

In addition, in continuation of my previous work regarding domain-specific sentiment analysis, I compared the performances between models employing different dictionaries:

1. General, binary sentiment-based model ( positive or negative words )
2. Moral-emotional sentiment-based model ( moral-emotional words such as peace and punish )
3. Outrage-fear sentiment-based model

Data

The scraped tweets were posted by all 100 U.S. Senators during the year leading up to the 2016 U.S. election: from November 2015 to October 2016. (n = 99,750)

Features

• The number of followers
• URL and Media Attachment
• Political ideology scores
• Sentiment Analysis
• Gender
• Social support

Target

The main task was to classify whether a tweet is retweetable or not. In defining what a retweetable message is, I referred to the median retweet counts in the dataset, which was 6. Thus, Twitter messages retweeted 6 times or more were classified as 'retweetable', and the other were categorized as 'not retweetable'.

Heatmap for the Correlation Matrix for All Features

Feature Importance Plot for the Best Model

importance type='weight'

Performance

: The XGBoost models using the outrage-fear dictionary rendered the highest AUC of 0.8393.