Twitter Bot Detection Through Unsupervised Machine Learning
Jeremy Wu, Eric Teng, Ziyue Cao · 2022 IEEE International Conference on Big Data (Big Data) · 2022
Identification a nd r emoval o f f ake internet accounts have become crucial for internet safety. Much of existing research for bot detection uses supervised machine learning, in addition to detecting URL usage and sentiment analysis. Due to the limited availability of labeled bot data and the difficulties s upervised a lgorithms f ace i n detecting adaptive behavior, an unsupervised approach would be more appropriate. Past studies have attempted this by clustering accounts based on post activity and detecting spam through embedded URLs. Our research attempts to detect a broader range of bots through K-Means and Agglomerative clustering using account activity, popularity, and Twitter verification, among others. After using the Scikit-learn Python library and collecting data from the Twitter API, we used a dataset of around 2,000 known bot and human accounts from the Bot Repository and 11,000 unlabelled accounts to create four distinct clusters. We then measured our results based on our confidence that an account in a particular cluster was a bot or a human.