Hate speech detection on social networks is a critical task for ensuring healthy and sustainable public conversation around the world. Social media platforms today are held accountable for the toxic content on their platforms, and machine learning assumes a critical role in detecting malevolent users at scale. The complexity of natural language makes this task hard to solve by purely from text. There has been some promising work on exploiting information from the topology of the social network to infer the toxicity of the user. Our work compares various approaches to this problem on a semi-supervised learning task where we have limited number of labeled user, and we build on the current Graph Neural Network methods by utilizing not only the topology of the graph, but also the evolution of the topology over time. Our results show that in the case where there are limited number of labeled users, algorithms that exploit the graph topology outperforms content-base approaches that use users' features and tweets. Additionally, we show that utilizing the dynamic evolution of the graph improves the performance of hate-speech prediction.