Authors:
Pranav Venkatesh、Kami Vinton、Dhiraj Murthy、Kellen Sharp、Akaash Kolluri
Paper:
https://arxiv.org/abs/2408.06900
Introduction
Social bots are automated accounts on social media platforms designed to mimic human behavior and engage with users. These bots can manipulate public perception, spread disinformation, and amplify fringe agendas, leading to significant societal harm. While mainstream platforms like Twitter have developed robust bot detection tools, niche and fringe platforms such as Parler, Gab, and Gettr remain vulnerable. To address this gap, the authors introduce Entendre, an open-access, scalable, and platform-agnostic bot detection framework.
Previous Work
Botometer is a notable example of a successful bot detection service for Twitter. It has been used to identify automated accounts behind various misinformation campaigns, including the 2016 Russian interference campaign. However, there is limited research on bot detection for fringe platforms like Parler. One study suggested that social bots might be rampant on Parler, exacerbating echo chamber effects and influencing user behavior.
Methods
Dataset
The dataset used to train the machine learning model was scraped from Parler, containing 183 million posts made by 4 million users between August 2018 and January 2021. A subset of 50,000 posts from verified social bot and human accounts was used for model training. The data was pre-processed to normalize features and impute missing information.
Machine Learning Model
The authors opted for a feature-based machine learning approach using a random forest classification model. This model was chosen for its effective classification of numerical inputs, performance accuracy, and popularity in social bot detection. Hyperparameter tuning was performed using sequential model-based optimization to improve the model’s performance.
Backend API
Entendre’s backend API, written in Express.js, retrieves account posts and metadata from Parler using an unofficial API. The collected information feeds into a supervised machine learning model, which computes and returns a bot-likelihood score for the account. The API is publicly accessible for direct bot detection.
Frontend Website
Entendre is also available through a public-facing web application created in React.js. Users can input a Parler account handle to receive a bot-likelihood score and relevant visuals for further analysis, such as time-based graphs of account activity and charts of frequently used terms/hashtags.
