Platforms that are based on heterogeneous architectures require an intelligent resource manager. An intelligent resource manager should be able to accurately predict the future workload of the system at hand and take it into consideration. In this paper, we show that there exist patterns in the interarrival times of resource requests, and that these patterns can be used for modeling and prediction of the future arrivals. To this end, we develop a two-phase machine-learning-based framework and apply it to real data. First, in the offline phase of our framework, the interarrival times are clustered based on a number of extracted features, and then an adequate modeling and prediction method is selected for each detected cluster. It is shown that, due to the intricate and varied nature of interarrival times, a universal modeling and prediction method does not provide optimal results, and a customized method should be applied to each of the detected clusters. Second, in the runtime phase of our framework, the results provided from the offline phase are used to perform computationally cheap prediction. The experimental results show that our approach has a prediction error below 12% and provides an error reduction of more than 17% in comparison with a straightforward method.