Speaker
Description
In particle physics, papers on weakly supervised searches for new physics have started accumulating numbers. This is the paradigm of trying to learn the differences between a data-driven background estimate and real data, in order to isolate some small contribution of anomalous data points - the signal - in the real data. The technique has been shown to work in simulation, and has been applied in several studies by experiments at the Large Hadron Collider. In recent years, boosted decision trees, or BDTs, have been shown to out-perform neural networks for binary classification in weakly supervised search contexts. They have several advantages; for example, training is faster, and they are more robust to noisy or uninformative features in the data.
However, BDTs also have some quirks. They work by partitioning the feature space of the data into varyingly sized, axis-aligned hyperboxes. This discretization can be considered unphysical, at least in the sense that physics is considered smooth down to the quantum level. Additionally, this partitioning is dependent on the complexity of the BDT as well as the amount and density of data. Combined, this has consequences for which data is actually considered "anomalous", in a way that may not be visible in high-level quantities like receiver operating curves and significance improvement characteristics.
In this talk, I dive down the rabbit hole of BDT anomaly detection as I've encountered it in my work studying high-energy physics data. I will visualize the choices a BDT makes in different conditions, which may or may not be in line with physics goals, and try to understand those choices in terms of statistical properties.