Tailored Audiences on Twitter

Insights from ~1,000,000 tweets about the 2016 Democratic National Convention

Outline

• Background
• Data
• Text Preprocessing
• Identifying. and Labeling Groups
  • Network Analysis
  • K Means Clustering
• Modeling
  • Selection
  • Scores

Process Overview

Background

Standard Audience Targeting:

• Age
• Gender
• Geography
• Generic interests (eg “shopping”, “sports”)

My Tailored Audiences:

• Mothers who are fans of Michelle Obama
• Men who are fathers and religious
• Teenagers who support Bernie Sanders
• etc...

Network Analysis

Methods

• 1.) Create a graph where points represent users and lines represent conversations (tweets)
• 2.) Nodes are grouped into communities based on the frequency of their interactions with other users (mentions, replies)
• 3.) For each community, extract the keywords that define it using TF-IDF
• 4.) Hand Label the communities

Labels

• Bernie supporters, progressives who are talking about a politicla revolution. Anti-war, justice and unity
• Women, Mothers, Black, talking about Michelle Obama
• Mainstreams anchors hournalists hosts, mostly at CNN
• Men, fathers, dads, professionals in news media who are conservatives and talk about hillary
• Mothers
• religious conservatives
• Journalists talking about the DNC chair debbie wazzerman schultz
• Hispanic and Latinx Voters
• Authors and writers
• News people
• Conservatives talking about donna brazille and implicating corruption, talking about wikileaks
• Religious males (husbands) who work in media (radio
• Muslim Voters
• Reporterrs
• Teenagers, Young Females
• Talking about mental illness and demi lovato (?)

Network Graph

KMeans Clustering

Methods

• 1.) Using a pre-trained word2vec model trained on 2 billion tweets (stanford), vectorize text data (aka turn the words into a numerical representation)

• ex: King - Man + Woman = Queen
  

• 2.) Use K-Means algorithm to group users that use similar language together
• 3.) Label each user with their assigned cluster
• 4.) Hand Label the clusters using topic modeling
• 5.) Reduce dimensions to vizualize the clusters in 2d

Labels

• Talking about Debbie Wasserman Schultz (unfavorably)
• Women discussing women’s issues
• Expressing concern about trump with the nuclear codes
• Talking about love, God, good vibes
• Watching the debate with friends, supporting Hillary Clinton
• Proud Americans
• Conservative sports fans
• Barack and Michelle Obama fans
• Barack and Michelle Obama fans
• Young people, Bernie Sanders supporters

Train and Validate Models:

Cluster Classifier Benchmarks

Community Classifier Benchmarks