Using a Semi-Evolutionary Algorithm to Optimize Deep Network Hyper-Parameters with an Application to Donor Detection

The performance of deep learning depends on finding an appropriate architecture to fit the model to the data. As deep networks (DNs) have been scaled up to different application areas, network architectures have become difficult to design manually. This paper proposes a new method, SEODP, for optimizing DN architectures and hyper-parameters by using a semi-evolutionary and semi-random approach, which can automatically evolve DN architectures for a given dataset. The experimental results show that, compared with manual search, SEODP is much more effective, and compared with grid search, SEODP can achieve optimal performance much faster. We also apply SEODP to a real-world social behavioral dataset from a charity organization for detection of potential donors. The results show that SEODP is a promising approach for optimizing DN architectures over different types of datasets.