Argumentation is the processes of producing and evaluating arguments based on context of a discussion, dialogue or conversation. In information-overloaded world, manually extracting argument relation is time-consuming, knowledge-intensive and prone bias. The entire goal of argument mining is automatic identification and extraction of arguments along with their relation from large unstructured data that are useful for reasoning engine and computational model. It has an important role in individual and group decisions, text summarization, business and governmental analysis and so on. There have been many efforts tried on argument relation prediction for different languages. However, most of the studies concentrated on English and other European languages. Argument component and relation classification(predictions) are two main subtasks in argument mining. In this study, we design and implement argument relation prediction for Amharic language using supervised machine learning. The proposed system consists, Preprocessing, for removing extraneous data and make the data for next component of learning algorithm. Feature generation, used to generate general features for propositional semantic similarity and discourse marker features. Finally, feature extraction used to extract specific features for argument relation prediction tasks. We have done nine experiments with three different supervised learning algorithms using 815 argumentative sentences to evaluate our argument relation prediction. The experiments are conduct using discourse marker, propositional semantic similarity and combined both discourse marker and propositional semantic similarity. All experiments are evaluated by Support Vector Machine(SVM), Multi-Layer Perceptron(MLP) and Naïve Bayes classifiers. From the experiments, we have obtained highest weighted averages of F-score 85%, 87%, 88% in experiment one, two and three respectively. From our experimental result we have observed that better performance is achieved by employing the combined features and the SVM classifier is proved to be a better learning classifiers for our Amharic argument relation prediction task.