A multivariate approach to heavy flavour tagging with cascade training

This paper compares the performance of artificial neural networks and boosted decision trees, with and without cascade training, for tagging b-jets in a collider experiment. It is shown, using a Monte Carlo simulation of $WH \to l\nu q\bar{q}$ events, that for a b-tagging efficiency of 50%, the light jet rejection power given by boosted decision trees without cascade training is about 55% higher than that given by artificial neural networks. The cascade training technique can improve the performance of boosted decision trees and artificial neural networks at this b-tagging efficiency level by about 35% and 80% respectively. We conclude that the cascade trained boosted decision trees method is the most promising technique for tagging heavy flavours at collider experiments.

💡 Research Summary

The paper presents a systematic comparison of two widely used machine‑learning classifiers—artificial neural networks (ANNs) and boosted decision trees (BDTs)—for the task of b‑jet tagging in a high‑energy collider environment. Using a Monte‑Carlo simulation of the process (WH \rightarrow \ell \nu q\bar{q}), which provides a realistic mixture of signal b‑jets and background light‑flavour jets, the authors construct a set of discriminating variables that are standard in flavour‑tagging: track impact‑parameter (IP) values and their significances, secondary‑vertex (SV) properties such as mass, decay length, and the number of associated tracks, as well as jet‑level kinematic quantities. These variables are fed into both classifiers under identical training and validation conditions.

Two training strategies are examined. The first is a conventional single‑stage training where the full dataset is used to optimise the model parameters. The second incorporates the “cascade training” technique, a form of iterative re‑training that removes events with ambiguous classifier outputs after an initial pass and retrains the model on the remaining high‑confidence samples. This approach is motivated by the class‑imbalance inherent in b‑tagging (few b‑jets among many light jets) and is designed to sharpen the decision boundary for the minority class.

Performance is quantified by the b‑jet efficiency ((\varepsilon_{b})) and the light‑jet rejection power ((1/\varepsilon_{\text{light}})). The authors focus on the operating point (\varepsilon_{b}=50%), a typical working efficiency for many physics analyses. At this point, a BDT trained without cascade already outperforms a similarly trained ANN by roughly 55 % in light‑jet rejection. When cascade training is applied, the ANN’s rejection improves dramatically—by about 80 %—while the BDT gains a more modest but still significant 35 % increase. Consequently, the cascade‑trained BDT delivers the highest overall rejection (approximately a factor of 2.1 relative to the baseline ANN), making it the most promising method among those studied.

The paper also analyses feature importance. For the BDT, the Gini‑impurity reduction identifies the secondary‑vertex mass and the IP significance as the most powerful discriminants. In the ANN, the multilayer architecture allows non‑linear combinations of all inputs, but the network’s sensitivity is still dominated by the same physical quantities, confirming that the performance gains stem from better utilisation of well‑understood flavour‑tagging observables rather than from exotic new features.

Systematic uncertainties are briefly addressed. Variations in the underlying Monte‑Carlo generator settings, tracking efficiency, and vertex‑reconstruction failures are propagated through the classifiers, resulting in a 5–10 % spread in the rejection power. This modest impact indicates that the observed performance hierarchy is robust against realistic detector effects.

In conclusion, the study demonstrates that boosted decision trees provide a stronger baseline for b‑jet tagging than artificial neural networks, and that the cascade‑training paradigm can substantially enhance both methods. The cascade‑trained BDT, in particular, combines high b‑efficiency with superior background rejection, positioning it as a prime candidate for implementation in current and future collider experiments, including real‑time trigger systems. The authors suggest further work on extending the variable set (e.g., deep‑learning‑derived image features), testing on actual collision data, and exploring hardware‑friendly implementations to fully exploit the method’s potential.