The traditional Monte Carlo event selection process does not have a well-defined method to quantify the effectiveness of the event selection.
In addition, the selection task normally produces a binary result only, either a good or bad fit to the event of interest. A bad fit is then assumed to be a different event type, and is removed from the analysis.
In a broader perspective, an unsupervised classification algorithm would offer the possibility to discover rare events which may not be expected or are overlooked. These events would likely be filtered out using the traditional methods. From a practical point of view, compared to supervised learning, it also avoids the necessary labeling task of the learning set events, which is error prone and time consuming.