Signal Enhancement Using Multivariate Classification Techniques and Physical Constraints

Abstract

We report on an empirical comparison of several multivariate classification techniques (e.g., random forests, Bayesian classification, support vector machines) for signal identification; our experiments use K * + mass as a test case. We show 1) the effect of using different cost matrices in generalization performance and 2) how information about physical constraints obtained from kinematic fitting procedures can be used to enrich the original feature representation. The latter step is done through a derivation of Λ particle parameters (e.g., momentum, energy, and mass) using kinematic fitting; the degree of fit using a χ 2 statistic is used as a new feature. Overall, our goal is to investigate how to incorporate physical constraints to improve classification performance.

Multivariate side-band subtraction using probabilistic event weights

“…If one has a priori knowledge of certain features of the signal and background, algorithms can be created to separate the two types of data. Procedures have been developed to handle many of these situations (see, e.g., [1][2][3]). Two of the more common methods for performing this type of data classification are neural networks and decision trees.…”