Unmasking Trees for Tabular Data
Authors: Calvin McCarter
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On a benchmark for out-of-the-box performance on 27 small tabular datasets, Unmasking Trees offers leading performance on imputation; state-of-the-art performance on generation given data with missingness; and competitive performance on vanilla generation given data without missingness. We showcase Unmasking Trees on two tabular case studies, and on the benchmark of 27 tabular datasets presented by Jolicoeur-Martineau et al. (2024b). On this benchmark for out-of-the-box performance on small tabular datasets, our approach offers leading performance on imputation and state-of-the-art performance on generation given data with missingness; and it has competitive performance on vanilla generation without missingness. We also demonstrate that Balto Bot is on its own a promising method for probabilistic prediction, showing its advantages on synthetic case studies and on a heavy-tailed sales forecasting benchmark. |
| Researcher Affiliation | Industry | Calvin Mc Carter EMAIL Big Hat Biosciences |
| Pseudocode | Yes | The training algorithm for Unmasking Trees is given in Algorithm 1. Algorithm 1 Unmasking Trees training Require: dataset X RN D; duplication factor K. ... The training and inference algorithms for Balto Bot are given in Algorithms 2 and 3, respectively. Algorithm 2 Balto Bot training Require: dataset (X RN D, y RN); Balto Bot meta-tree height H; ... Algorithm 3 Balto Bot inference Require: input query x RD; trained Balto Bot model. |
| Open Source Code | Yes | Finally, we provide code with an easy-to-use sklearn-style API at https://github.com/calvinmccarter/ unmasking-trees. |
| Open Datasets | Yes | We showcase Unmasking Trees on two tabular case studies, and on the benchmark of 27 tabular datasets presented by Jolicoeur-Martineau et al. (2024b). On this benchmark for out-of-the-box performance on small tabular datasets... Iris dataset (Fisher, 1936)... We employ the M5 sales forecasting Kaggle dataset (Makridakis and Howard, 2020) to compare Balto Bot with other probabilistic prediction methods. |
| Dataset Splits | Yes | The dataset has five years of sales data from ten Walmart stores, and the task requires predicting the (heavy-tailed) number of units sold given a product s attributes and previous sales. We use the exact same data preparation used for Treeffuser (Beltran-Velez et al., 2024) experiments, which yields 1k products, 120k training samples, and 10k test samples. ... For comparability, for this (and only this) dataset, we also evaluate Balto Bot with hyperparameter tuning, using the same setup used for all other methods (10 folds, each with 80%-20% train-validation split, and 25 Bayesian optimization iterations). |
| Hardware Specification | Yes | Experiments were performed on a i Mac (21.5-inch, Late 2015) with 2.8GHz Intel Core i5 processor and 16GB memory. |
| Software Dependencies | No | The paper mentions using XGBoost models and an sklearn-style API, but no specific versions for these or other software dependencies are provided. For example, it mentions "XGBoost hyperparameters were set to their defaults" but not the version of XGBoost used. |
| Experiment Setup | Yes | Results were obtained always using our method s default hyperparameters: Balto Bot tree height of 4, and duplication factor K = 50. These hyperparameter values were tuned on the Two Moons and Iris case studies, then applied without further tuning to the remaining experiments... XGBoost hyperparameters were set to their defaults. ... We optimize over the following XGBoost hyperparameter spaces: learning_rate log-uniform(0.05, 0.5), max_leaves {0, 25, 50}, and subsample log-uniform(0.3, 1). |