MuHBoost: Multi-Label Boosting For Practical Longitudinal Human Behavior Modeling
Authors: Nguyen Thach, Patrick Habecker, Anika Eisenbraun, W. Alex Mason, Kimberly Tyler, Bilal Khan, Hau Chan
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments to evaluate Mu HBoost and its variants on 13 health and well-being prediction tasks defined from four realistic ubiquitous health datasets. Our results show that our three developed methods outperform all considered baselines across three standard MLC metrics, demonstrating their effectiveness while ensuring resource efficiency. |
| Researcher Affiliation | Academia | Correspondence to EMAIL. University of Nebraska-Lincoln. Lehigh University. |
| Pseudocode | Yes | Algorithm 1 Cluster Sampling (Extended) ... Algorithm 2 Mu HBoost[CC] |
| Open Source Code | Yes | We refer readers to Sections 4.1 and 4.2 as well as Appendices C.1 and C.2 for complete details on reproducing our results, including the link to our anonymous Git Hub repository10. ... 10https://github.com/Anony Mouse3005/Mu HBoost |
| Open Datasets | Yes | We consider four ubiquitous health datasets that contain both time-series and auxiliary data as well as sufficient information to form multiple labels for each: Life Snaps (Yfantidou et al., 2022), GLOBEM (Xu et al., 2022), college students (Co St) (Hayat & Hasan, 2023), and PWUD (Tyler et al., 2024). The first two are publicly available and have previously been considered by Englhardt et al. (2024); Kim et al. (2024) (as mentioned in Section 2), whereas the latter two are novel and require submitting an IRB protocol and ethical research plan to their authors. |
| Dataset Splits | Yes | For each considered set of experimental configurations, we used the split ratio of 50/10/40 (for train/validation/test set), which follows Summary Boost s evaluation, for a total of 10 different splits. For partitioning multi-label datasets into training (train+validation) and test sets, we adopt the iterative stratification algorithm13 (Sechidis et al., 2011). |
| Hardware Specification | Yes | All experiments were conducted under Ubuntu 20.04 on a Linux virtual machine equipped with NVIDIA GeForce RTX 3050 Ti GPU and 12th Gen Intel(R) Core(TM) i7-12700H CPU @ 2.3GHz. |
| Software Dependencies | Yes | We used PyTorch 1.13, CUDA 11.7, Open AI 1.23, and scikit-learn 1.3. |
| Experiment Setup | Yes | The number of boosting rounds T is set to a large value of 100 for training till convergence (typically within 10 20 rounds) and the size of the representative subset s is set as large as possible11 (without exceeding the maximum context length) to 10. We set µ, the nonzero hyperparameter for raising the bar for each weak learner, to 0 since we notice no significant increases in predictive performance otherwise. The stopping threshold at each round for both Mu HBoost and Mu HBoost[LP+], defined by 1 1/K µ, is hence 1 1/ min{N, 2Q} (below which is considered satisfactory and no further resampling is needed in accordance with Summary Boost). For Mu HBoost[CC] (K = 2), we introduce the discount factor γ = 0.95 into this threshold, which becomes 1 (1/2)γq for each label q [0, Q 1], to relax the training further down the chain (i.e., subject to error propagation). |