https://jjcit.org/paper/187 10.5455/jjcit.71-1667394363 Kennedy Malanga Ndenga Hate speech detection,TensorFlow decision forests,Gradient boosted trees,Universal sentence embedding,National cohesion and integration commission 4 4427 1333 2-Nov.-2022 4-Feb.-2023 14-Feb.-2023 Detecting and controlling propagation of hate speech over social-media platforms is a challenge. This problem is exacerbated by extremely fast flow, readily available audience and relative permanence of information on social media. The objective of this research is to propose a model that could be used to detect political hate speech that is propagated through social-media platforms in Kenya. Using Twitter textual data and Keras TensorFlow Decision Forests (TF-DF), three models were developed; i.e., Gradient Boosted Trees with Universal Sentence Embedding (USE), Gradient Boosted Trees and Random Forest, respectively. The Gradient Boosted Trees with USE model exhibited a superior performance with an accuracy of 98.86%, a recall of 0.9587, a precision of 0.9831 and an AUC of 0.9984. Therefore, this model can be utilized for detecting hate speech on social media platforms.