Seasonal time series exploration using conditional probabilistic graphical approach

Model-Based Machine Learning is an approach in which every assumption related to the problem domain is explicitly represented. Probabilistic Graphical Model (PGM) is one of its applications combining probabilistic and graphical approach. PGM application in sequential prediction problems is useful to keep the prediction in the track. A conditional version of PGM, Continuous Conditional Random Fields is discussed in this work. CCRF boosts the baseline predictor(s) that can be any conventional machine learning, such as artificial neural network, tree, and other predictors or regressor. CCRF can capture the relationship between individual predictions made by the baseline predictor(s) at different steps, yet those steps need not necessarily to be consecutive to each other. It allows CCRF to explore the time series sequence characteristic even much deeper, especially if the time series is seasonal. The experiment result show that CCRF can proportionally alleviate the error rate of its baseline by improving the baseline up to 54%.