The number of customers is an important indicator for companies to know the success of a product and service offered. In general, customers are grouped into two categories, loyal customers and disloyal customers. This disloyal customer refers to customers who have stopped using a product and service from a company or are often referred to as churn. For that, we need a system that can predict whether the customer has the potential to experience churn or not. The system is required to predict well with one of the indicators, namely a lowcost value. One way to reduce the cost value is to use the optimizer function. Researchers use deep learning algorithms to create predictive models. A total of 3333 rows and 20 columns of telecommunication customer public data are used as datasets. This study also compares several optimizer algorithms to find the lowest cost value. In addition to the cost value, the accuracy and f1 score results are also used as other considerations. Researchers also use cross-validation for the training process and validating the model created. To deal with the imbalance class dataset, the researcher uses the unweighted method on the cost function. The evaluation results show that SGD has the lowest cost value, which is 0.261. Meanwhile, the AdaGrad matrix classification shows the best value for accuracy and f1 score with 0.926 and 0.846, respectively