Sosyal ağlarda sanal zorbalığın otomatik olarak tespit edilmesi

Cyberbullying has become a major problem that affects children and youngsters especially. In this thesis, it is aimed to detect cyberbullying content in social networks automatically. In this direction, a comprehensive dataset, which includes social media features (e.g., number of the sender followers), was systematically prepared. Then, the characteristics of cyberbullying are inspected by analyzing the natural language processing features (e.g., the number of title words) and social media features on the prepared dataset. It is seen that some of the social media features are strongly related to cyberbullying. Additionally, some association rules between social media features and cyberbullying were captured, such as users that have more followers on social networks are disinclined to post online bullying content. The obtained results show that social media features would be promising in automatically detecting harmful content in social networks. Accordingly, machine learning algorithms experimented on two different variants of the prepared datasets. The first variant includes only textual features, whereas the second variant consists of the determined social media features and textual features. It is observed that each experimented machine learning algorithm gives more successful prediction performance on the variant containing social media features. Further experiments in machine learning were conducted by implementing word embedding approaches in the feature extraction to increase the performance of the applied machine learning algorithms. Lastly, an open web service that uses the trained machine learning models for cyberbullying detection was published to motivate programmers to develop real-time applications without studying or knowing the machine learning process.