{ "fetched_at": "2026-10-19T11:37:50Z", "source_url": "https://www.kaggle.com/competitions/2024-b2sw-ml/data", "pages": [ { "name": "data-description", "content": "The data used in this exercise consists of charged final-state particles (\\\\(e, \\mu, \\pi, K, p\\\\)) reconstructed in \\\\( B^{0} \\rightarrow D^{\\ast\\pm} \\ell^{\\mp} \\nu \\\\) signal MC. There are about 600,000 particles in the dataset, with the majority being hadrons (~470,000), and almost equal amounts of \\\\(e\\\\) and \\\\(\\mu\\\\) (~66,000 each) for the rest.\n\nOur task is to identify \\\\(e\\\\), \\\\(\\mu\\\\), and hadrons using detector information from Belle II. The target variable is derived from a truth-matching variable where different hadron species are grouped, resulting in three classes in the dataset.\n\nThe dataset is split into `train.csv` and `test.csv` files with a 70%-30% split. Training features are provided in both files, but only `train.csv` contains the target variable. Your job is to train and optimize a model using `train.csv` and make predictions on the `test.csv`.\n\nAfter making predictions, you can upload them to the leaderboard. The leaderboard will use 30% of the data in the `test.csv` to evaluate the performance of your model and display the results publicly. When the competition is over, the leaderboard will re-evaluate models using the remaining 70% of the data in `test.csv` to announce the winner.\n\n## Files\n\n* **train.csv** - the training set\n* **test.csv** - the test set\n* **example_submission.csv** - an example submission file in the correct format\n\n## Columns\n\n* `index` - particle index\n* `features`/`training_variables` - kinematics, ECL cluster variables, subdetector PID likelihood\n* `target` - the class label\n* etc.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "In this hands-on exercise, you are encouraged to use Auto Machine Learning libraries that automate the end-to-end process of applying machine learning to real-world problems. These libraries handle tasks such as data preprocessing, feature selection (!!), model selection, and hyperparameter tuning. This greatly simplifies the technical details, allowing us to focus more on physics. The goal of this challenge is to identify leptons (electrons and muons) from other particles using simulated data (MC) from the Belle II experiment.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Evaluation", "content": "Your ML model(s) should be submitted and evaluated on the leaderboard. Submissions are evaluated on [F1-score](https://en.wikipedia.org/wiki/F-score) between the predicted probability and the observed target. This metric measures the ability of the model to discriminate between the classes. A higher F1 score indicates better performance.\n\nThe public leaderboard will display your model performance. Each team (up to 3 people) can have 5 submissions per day and choose up to 2 models from all your submissions. The winner will be announced on Friday based on the private leaderboard.\n\n## Submission File\nThe submission format will be a CSV file with two columns. The first column is the `Id` column, which matches the IDs of events in the test. The second column is the predicted class label by your model.\nThe file should contain a header and have the following format:\n\n index,predict\n 2,0\n 5,0\n 6,0\n etc.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "rules", "content": "- Participants must register for the competition through Kaggle.\n- Each participant can make up to 5 submissions per day.\n- Each participant chooses up to 2 submissions for the final private leaderboard.\n- The competition is open to individuals and teams (up to 3 people per team).\n- Sharing of solutions is prohibited until the competition ends.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "abstract", "content": "In high-energy physics experiments, collisions between particles produce a variety of secondary particles. Identifying these particles accurately is vital for understanding the underlying physics. PID helps researchers distinguish between different types of particles (e.g., electrons, muons, pions, kaons, and protons) based on their physical properties, enabling precise measurements and the validation of theoretical models.\n\nAt Belle II, PID is achieved through a combination of sophisticated subdetectors and software algorithms. Key subdetectors involved in PID include:\n\n- Central Drift Chamber (CDC): Measures the energy loss \\\\(dE/dx\\\\) due to ionization, important for identifying particles with low momentum.\n- Cherenkov Detectors (TOP and ARICH): Utilize the Cherenkov effect to identify particles based on the angle and intensity of emitted Cherenkov light.\n- Electromagnetic Calorimeter (ECL): Measures the energy deposited by particles, crucial for identifying electrons and photons.\n- K-long and Muon Detector (KLM): Identifies muons by measuring their penetration depth and interaction characteristics in the detector layers.\n\nThese hardware components provide raw data, which is then processed using advanced software algorithms to classify particles. Machine learning techniques play an increasingly important role in improving the accuracy and efficiency of these PID algorithms.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "B2SW ML hands-on series", "content": "* Continuum Suppression at Duke 2023\nKaggle: https://www.kaggle.com/competitions/cshandson/\nIndico: https://indico.belle2.org/event/8841/contributions/61156/", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }