Customer churn is a major issue for operators that may greatly affect theirrnrevenue. Even if in Ethiopiaâ€™s monopolistic market, there is still a challenge inrnthe form of subscriber service termination. To solve this issue, operators mustrnfirst identify potential service termination in advance and then take proactivernmeasures to reduce the number of terminations. For this purpose, telecomrnoperators need a prediction model to predict correctly and timely potentialrnservice termination subscribers based on collected service usage-related data.rnTo anticipate subscriber behavior, the performance of several prediction methodsrnhas been explored. However, the performance of such algorithms is notrnexamined in the context of Ethiopia postpaid mobile service and in the case ofrnmulti-class, where ethio telecom possesses multi-class data that enable developrna multi-class prediction model for mobile subscriber service termination.rnThe goal of this study is to investigate the performance of prediction learningrnalgorithms with multi-class scenarios for predicting service termination inrnthe context of Ethiopian postpaid mobile subscribers. The algorithms investigatedrninclude J48 Decision tree, Random Forest (RF), and Multilayer Perceptronrn(MLP), and their performance is measured in terms of prediction accuracy,rnprecision, recall, and F-Measure. Cross-validation (k=10) techniques andrna multi-class dataset are used to test the performance of algorithms. WEKA 3.9.4rntool algorithm implementation was utilized for performance evaluation. As arnresult, the J48 and RF prediction algorithms almost have the same performancernon all performance parameters result. However, MLP algorithm achieved a lowperformancernscore compared to J48 and RF, and the accuracy of J48, RF, andrnMLP are 94.9%, 95.1%, and 93.3% respectively.