Parameterized and Custom Loss Functions: A Case for Network Traffic Prediction

We show the importance of designing and using application-specific, parameterized, custom loss functions in predictive Deep Neural Network (DNN) applications. We consider a cellular traffic prediction application in next-generation communication systems where predicting less than the actual traffic value is more harmful than predicting more than the actual traffic value. Cellular network operators aim to maintain service-level agreements (SLAs) that carry high penalty costs for violations, i.e., negative prediction errors. Conversely, the cost of overprovisioning, i.e., positive prediction errors, is less significant to operators compared to SLA violations. Standard loss functions, being symmetric, cannot account for the differing importance of positive and negative prediction errors. In this work, we introduce and compare several custom loss functions where the weight of the negative prediction error is determined by the SLA violation rate. We find that our proposed weighted Mean Exponential Error (wMEE) based loss function results in the smallest overprovisioning volume while adhering to a given SLA violation rate. More crucially, our findings suggest that custom loss function parameter design should be incorporated into the hyperparameter tuning process for DNN applications.