Profile Identification on Social Media Using Artificial Intelligence Techniques for Textual Similarity

Abstract

Identifying user profiles that refer to the same individual is a critical task in applications such as brand protection, behavioral analysis, and digital forensics on social media. Despite their importance, most existing methods are limited when confronted with subtle variations in usernames, such as insertions, deletions, or character substitutions, which are especially common on platforms such as Instagram, where usernames often serve as the sole public identifier. In this paper, we address the challenge of matching user profiles on Instagram by exploring artificial intelligence techniques for textual similarity, focusing on subtle variations in usernames that often hinder traditional methods. We propose and evaluate two complementary strategies: a lexical approach using Levenshtein distance with a dynamic threshold and a semantic approach based on RoBERTa-base fine-tuned for binary classification. Experiments on 48,609 unique username pairs derived from over 318,000 real-world Instagram accounts revealed that the AI-based semantic model significantly outperformed the lexical baseline, achieving 92% accuracy and an F1-Score of 89%, with a substantial reduction in misclassification rates. The results demonstrate that transformer-based models can capture latent patterns beyond surface-level similarity, highlighting their potential for profile identification on social media platforms and motivating further research on real-world identity resolution applications.