Streaming Hamiltonian Monte Carlo with Smooth Data Drift Adaptation

We build on the Streaming Hamiltonian Monte Carlo (SHMC) online machine learning method developed in the previous work. We further explore the problem of concept drift adaptation for Hamiltonian Monte Carlo, demonstrating the method's capability against several types of drifts known in the literature. In particular, we test against 'Sudden Drift”, “Gradual Drift”, “Incremental Drift”, and “Reoccurring Concepts” scenarios. We introduce an enhancement to our method to make the drift adaptation behavior tunable. The enhancement consists of parameterized exponential smoothing of the sampler parameter updates. Thus, we can reduce the weight of update in SHMC without compromising the model convergence. We further demonstrate the benefits of the enhancement with a simple nowcasting model on an open real-world dataset from the Energy Engineering domain. The data set consists of records from 15 years of power outages across the United States. SHMC model with smoothing results in better error metrics and less overfit than the baseline SHMC model.