Machine Learning for Disinformation Analysis

Hashtag Similarity Mapping via Dimensionality Reduction

​

​

Michael Rossetti

https://github.com/s2t2/ml-project-2022

Background

Background

The First Trump Impeachment

We have previously collected 67.7M tweets about the First Trump Impeachment (including 3.6M user profiles), and stored them in a Google BigQuery database for further analysis.

Previous Research

Bot Classification

Previous research goals:

• classify the likelihood any given user is a "bot" (i.e. an automated account)
• quantify the influence of the bots

​

"Bots, Disinformation, and the First Trump Impeachment"

- Michael Rossetti, Tauhid Zaman (https://arxiv.org/abs/2204.08915)

​

Previous Research

Sentiment Classification

Previous research goal: predict the degree to which any given tweet expresses either Pro-Trump or Anti-Trump sentiments

0

1

0.5

Explore results on our website

Problem

The Problem

Evidence of Disinformation Spread

The Problem

What is Q-Anon?

Q-Anon is a popular conspiracy theory / disinformation campaign. Q-Anon supporters use hashtags such as #QANON, #WWG1WGA, #GREATAWAKENING, etc. to communicate and organize. We saw Q-Anon supporters participate in the attack on the US Capitol on January 6, 2021. How can we prevent events like this in the future?

​

The Problem

Challenges facing Social Networks and Online Platforms

Challenges in Trust and Safety:

1. How to conduct near real-time monitoring and moderation of disinformation and other prohibited content?
2. How to keep up with malicious actors who evolve their tactics (e.g. using new hashtags over time)
3. How to reduce costs associated with this content moderation effort?

Related Work

Hashtag Co-occurrence Analysis

"A network-based approach to QAnon user dynamics and topic diversity during the COVID-19 infodemic"

- Wentao Xu, Kazutoshi Sasahara (https://arxiv.org/abs/2111.00537)

hashtag co-occurrence network hashtag semantic map

Research Plan

Objectives and Constraints

Goals / Objectives:

• Produce a two-dimensional mapping of the top hashtags from our Impeachment 2020 dataset
• Visually compare the distances between hashtags on the map to determine which topics are closest (i.e. most similar) to Q-Anon related terms
• Cluster the hashtags into more formal groups (STRETCH GOAL)

Assumptions / Constraints:

• Use unsupervised machine learning methods only
• Focus on the most frequently used hashtags only (because they are the most influential, and because we can only see so many points on a plot)

Research Plan

Our Contributions

Differentiating factors for this project:

• Related work uses hashtags in tweets, but we use hashtags in user profiles because we believe they provide a stronger signal about a user's sentiments
• Our dataset is unique, and focused on a domestic political event
• Related work constructs and utilizes a hashtag co-occurrence network, while our approach is simpler and does not require this network graph compilation step
• Related work focuses on hashtags co-occurring with "#QANON", but we focus on a broader range of top (i.e. most frequently used) hashtags, not just disinformation terms

Methods

Step 1: Data Collection

Twitter API Stream Listener

• Software written in Python (open source)
• Deployed to run continuously on a Heroku server
• Fetches data from the Twitter API
• Listens for any tweets mentioning Impeachment-related terms (see appendix)
• Stores tweet and user profile data in a Google BigQuery database

Step 2: One Hot Encoding

one-hot encodings

BigQuery database records

Step 3: Dimensionality Reduction

reduced "embeddings"

one-hot encodings

Step 4: Plotting the Embeddings

reduced "embeddings"

Step 5: Clustering the Embeddings

reduced "embeddings"

Results / Demo!

Colab Notebook

https://s2t2.github.io/ml-project-2022/

Dimensionality Reduction Techniques

• Principal Component Analysis (PCA)
• Benefits: interpretable / explainable
• Drawbacks: affected by outliers
• T-distributed Stochastic Neighbor Embedding (T-SNE)
• Benefits: more robust to outliers
• Drawbacks: less interpretable / explainable
• Uniform Manifold Approximation & Projection (UMAP)
• Benefits: fast, scalable, may perform better on complex data
• Drawbacks: less interpretable / explainable
• etc.

Dimensionality Reduction:

Principal Components Analysis (PCA)

Dimensionality Reduction:

T-distributed Stochastic Neighbor Embedding (T-SNE)

Dimensionality Reduction:

Uniform Manifold Approximation & Projection (UMAP)

Dimensionality Reduction:

Uniform Manifold Approximation & Projection (UMAP)

Dimensionality Reduction:

Uniform Manifold Approximation & Projection (UMAP)

Dimensionality Reduction:

Uniform Manifold Approximation & Projection (UMAP)

Why stop in two dimensions when we can try three?

​

​

Let's explore via live demo!

​

​

​

​

​

​

​

UMAP Enhanced Clustering:

Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN)

Clustering of the embeddings helps us identify lesser known terms

​

… which can now be monitored

UMAP Enhanced Clustering:

Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN)

Conclusions

• These methods can be used by social network trust and safety teams, and applied to any online discussion where hashtags are used
• Dimensionality reduction, when applied to a matrix of hashtags and one hot encoded user identifiers, can detect which terms are related to each other, including which terms are related to known disinformation terms
• These methods, due to their unsupervised nature, can detect new topics that may emerge over time, where human annotations may not yet be available
• The unsupervised approach also saves time and costs related to human labeling

Thank you!

Extra Slides / Appendix

Contact

Presentation by: Michael Rossetti

Contact: Email | LinkedIn | GitHub

Interests:

• Machine Learning (AI/ML)
• Natural Language Processing (NLP)
• Sentiment Analysis
• Social and Information Networks

​

Data Collection

Impeachment Topics

FYI - the term "rain" would also catch mentions of "#RainBows" (but not vice-versa)

Data Collection

Impeachment Timeline

We collected tweets from December 12, 2019 to March 24, 2020.

This covers the full Senate trial, including the time leading up and the time after.

Data Collection

Tweet Collection Timeline and Results

In total, we collected 67.6M tweets authored by 3.6M unique users.

Data Analysis

Inclusion of Hashtags in User Profiles

Of the 3.6M users, 451K users (12.5%) included at least one hashtag in their profile. They included 2.85 unique tags on average, for a total of 1.28M records.

​

We are most interested in the most frequently used hashtags, so we will focus on the top 25-75 hashtag range, with less than 200K users using those top tags.

​

Data Analysis

Inclusion of Hashtags in User Profiles

Of the 3.6M users, 451K users (12.5%) included at least one hashtag in their profile. They included 2.85 unique tags on average, for a total of 1.28M records.

​

We are most interested in the most frequently used hashtags, so we will focus on the top 25-75 hashtag range, with less than 200K users using those top tags.

​

Data Analysis

Top Hashtags found in User Profiles

Previous Analysis

Top Hashtags in Bot Profiles

Pro-Trump bots spread Q-Anon related content (via their User profiles)

Previous Analysis

Top Hashtags in Bot Tweets

Pro-Trump bots spread Q-Anon related content (via their Tweets)

Previous Sentiment Labeling Approach

Top Hashtags used by Anti-Trump Bot Cluster

Previous Sentiment Labeling Approach

Top Hashtags used by Pro-Trump Bot Cluster