Learning to rank Higgs boson candidates

Area of Science:

• High Energy Physics
• Particle Physics
• Machine Learning in Physics

Background:

• Precise measurement of the Higgs boson is crucial for discovering new physics.
• Machine learning is increasingly applied to analyze complex particle physics data, such as Higgs production via vector-boson fusion.

Purpose of the Study:

• To propose and evaluate learning to rank algorithms for Higgs boson event analysis.
• To compare the performance of pairwise ranking models against traditional classification methods.

Main Methods:

• Utilizing pairwise comparisons of signal and background events for training machine learning models.
• Developing a pairwise neural network algorithm combining convolutional neural networks (CNNs) and DirectRanker.
• Comparing the proposed algorithm with standard methods like CNNs, multilayer perceptrons (MLPs), and boosted decision trees (BDTs).
• Applying transfer learning techniques to enhance performance across different data types.

Main Results:

• Pairwise ranking models effectively increase training data through quadratic combinations.
• The proposed approach demonstrates robustness in unbalanced datasets and improves performance over pointwise models.
• The pairwise neural network algorithm shows competitive or superior performance compared to BDTs, CNNs, and MLPs in Higgs production channels.

Conclusions:

• Learning to rank offers a powerful alternative to classification for Higgs boson event analysis.
• Pairwise approaches enhance data efficiency and robustness, leading to improved performance in particle physics machine learning applications.
• Transfer learning further optimizes the performance of these models across diverse datasets.