In traditional machine learning, the goal is to “predict” the output of a function on previously unseen input data. The input data need not necessarily have a temporal dimension and the term “prediction” refers to the output of the learned function on a new data point. On the other hand, the task of forecasting is to predict the temporally future output of some function or the occurrence of an event. Time is thus a crucial component for forecasting. What makes forecasting particularly challenging is the fact that, from the (current) timepoint where a forecast is produced until the (future) timepoint for which we try to make a forecast, no data is available.
In particular, complex event forecasting assumes that we already have a pattern that we need to detect on event streams and that we need to then build a probabilistic model in order to forecasts occurrences of the pattern as well. For this thesis, the candidate will address the issue of unsupervised complex event forecasting, i.e., the situation where the pattern is unknown and the only available information concerns the labels of pattern occurrences. Another additional challenge is that current forecasting models are constructed based on “manually” provided features. The goal is to automate this process in order to discover the most important features through dimensionality reduction, feature extraction and feature selection techniques.
Bibliography:
Deliverable (besides the thesis itself): Publication quality report, 12 double-column pages