Teorija mrtvega interneta

Social media platforms play a central role in shaping digital identity and online communication, but they are increasingly affected by automated accounts and synthetic activity.Building on the Dead Internet Theory (DIT), which argues that much online interaction is artificially generated, this thesis investigates bot detection on Instagram, a platform that has received comparatively little academic attention despite its global relevance.The research applies the Design Science Research (DSR) framework in combination with the CRISP-DM process model to ensure both methodological rigor and practical relevance.Data was collected through automated scraping and incorporates three modalities: profile metrics, textual comments, and visual features.Following Georges' identity layer model, these were mapped to declarative, active, and calculated aspects of digital identity.Due to the scarcity of labeled datasets, a semi-supervised learning approach was adopted, leveraging a small labeled set alongside a larger pool of unlabeled accounts.Several machine learning models were trained and evaluated, including Random Forest, Support Vector Machines, Logistic Regression, and LightGBM.Results show strong performance for models trained on profile metrics, with Random Forest achieving 97% accuracy.In contrast, models using visual features performed significantly worse.When applied to the larger unlabeled dataset, Random Forest labeled around 2% of accounts as bots, while LightGBM identified nearly 18%, reflecting different sensitivities and assumptions in the models.The semi-supervised approach expanded the training dataset from 634 to 9,562 labeled instances, labeling 89.3% of the unlabeled data with high confidence.This demonstrates the method's effectiveness in low-labeled-data environments common to Instagram research.This thesis contributes by demonstrating the potential of multimodal, semi-supervised approaches for bot detection on Instagram.It also highlights methodological challenges, such as limited labeled data and model interpretability, while outlining directions for future research, including neural networks, sentiment analysis, and reverse image verification.