The CB-CTT university timetabling is an NP-Hard combinatorial problem that has been studied since 2007. However, it only has a small benchmark with few instances. New research on complex methods proposed plenty of new instances generated by a rigorous process. These instances simulate real-world characteristics. However, almost 30% of them are impossible to solve. In this paper, we propose a model that efficiently predicts their feasibility using a random forest. Moreover, an error analysis is conducted to understand the behavior of these instances. One predictor is created for each set of instances, artificial, and real. An analysis of these two models offers to compare behaviors considering the feasibility between the real and generated ones.
This work shows that new artificial instances present differences from the real world and are not useful to predict the feasibility of the other set. This is why we trained an instance-selection model that chooses a subset of artificial instances. These instances can be used to predict the feasibility of the real-world contrary to the initial whole set.