Grey Wolf Optimizer-Based Feature Selection with Mutual Information Refinement for Machine Learning-Based Bot Detection on Social Media

In the vast, noisy landscape of social media, a quiet but significant problem has taken root: the rise of automated accounts, or "bots." These are software programs designed to mimic human users, posting messages, liking content, and following others to spread everything from harmless news updates to dangerous misinformation and spam. While some of these automated accounts serve legitimate purposes, many are malicious, capable of swaying public opinion and undermining trust in online communities. For years, researchers have tried to build systems that can tell the difference between a real person and a robot. The challenge lies in the sheer volume of data; social media platforms generate millions of user profiles, each with dozens of different characteristics, from how many tweets they send to how long their username is. Sorting through this information to find the few clues that actually matter is like trying to find a specific needle in a haystack, but the haystack is constantly growing and changing.

A team of researchers at Dayalbagh Educational Institute in India has developed a new method to solve this sorting problem, aiming to make bot detection faster and more accurate. Instead of feeding every single piece of information about a user into a computer program, they created a two-step process to strip away the noise and keep only the most telling details. Their approach combines a search strategy inspired by nature with a statistical method that measures how much one piece of information tells us about another. By using this hybrid technique, they were able to reduce a massive list of sixty-eight different user traits down to a much smaller, more powerful group of just twenty-five, and eventually to an even tighter selection of ten features that proved to be the most effective for identification.

The researchers began with a large dataset containing information on 8,386 Twitter accounts, half of which were known bots and half were genuine humans. Each account was described by sixty-eight different features, such as the number of followers, the frequency of retweets, and the length of the user's bio. To find the best combination of these traits, the team first employed a method called the Grey Wolf Optimizer. This is a computer algorithm that mimics the social hierarchy and hunting behavior of grey wolves. In the digital version, a group of "wolves" explores the vast space of possible feature combinations, with the best solutions leading the pack toward the most promising area. This process acts as a filter, narrowing the sixty-eight original traits down to a more manageable set of twenty-five that showed the most promise for distinguishing bots from humans.

Once this initial group of twenty-five features was identified, the researchers applied a second layer of refinement using a concept called Mutual Information. Think of this as a way to measure how closely two things are related; in this case, it measured how much knowing a specific user trait helped predict whether that user was a bot. By ranking the twenty-five features based on this relationship, the team could see which ones were truly essential and which were merely redundant. They tested different group sizes, from the top four features up to the top twenty, to see how many were needed for the best results. This step ensured that the final list of features was not just small, but also packed with the most relevant information, removing any remaining clutter that might confuse the detection system.

To see if this streamlined approach actually worked, the researchers tested it against five different types of machine learning models, which are computer programs trained to recognize patterns. They used standard tools like decision trees and support vector machines, but also included a powerful model known as XGBoost. The results showed that by using their refined list of features, the systems performed exceptionally well. The XGBoost model, in particular, achieved a classification accuracy of 98.87% when using a subset of just eight to twelve features. This means the system correctly identified whether an account was a bot or a human in nearly every single case. Even with a very small number of features, the system maintained high performance, proving that it is possible to detect these automated accounts without needing to analyze every single detail of a user's profile.

The study also compared their new method against older techniques that used different search strategies, such as genetic algorithms and particle swarm optimization. While those older methods were able to find good solutions, the Grey Wolf Optimizer proved to be more efficient, finding a better set of features while reducing the total number of traits by over 63 percent. This reduction is significant because it means the detection system can run much faster and require less computing power, making it more practical for real-world use on large social media platforms. The researchers found that while some models struggled when given too few features, others, like the decision tree and XGBoost, thrived on the compact, high-quality data provided by their new framework.

Ultimately, this work demonstrates that effective bot detection does not require sifting through mountains of data. By carefully selecting the right clues, researchers can build systems that are both faster and more accurate. The study confirms that a combination of a nature-inspired search method and a statistical ranking tool can successfully isolate the specific behaviors that define a social media bot. While the researchers note that future work will need to address the constantly evolving tactics of bot creators and potentially incorporate more complex data types, their current findings offer a clear path forward. They have shown that with the right approach, it is possible to cut through the digital noise and identify the automated accounts that threaten the integrity of online communities.