Can Label-Specific Features Help Partial-Label Learning?
Authors: Ruo-Jing Dong, Jun-Yi Hang, Tong Wei, Min-Ling Zhang
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on both synthesized and real-world datasets are conducted and the results show that our method consistently outperforms eight baselines. |
| Researcher Affiliation | Academia | 1 School of Computer Science and Engineering, Southeast University, Nanjing 210096, China 2 Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, China 3 Computer Experimental Teaching Center of Southeast University {dongrj, hangjy, weit, zhangml}@seu.edu.cn |
| Pseudocode | Yes | Algorithm 1: The label-specific features approach in UCL |
| Open Source Code | Yes | Our code is released at https://github.com/meteoseeker/UCL |
| Open Datasets | Yes | Table 1 summarizes the characteristics of five commonly used UCI datasets... In addition to five synthetic datasets, we further test our method on five real-world partial-labeled datasets which are collected from various task domains, i.e., Lost (Cour, Sapp, and Taskar 2011), MSRCv2 (Liu and Dietterich 2012), Mirflickr (Huiskes and Lew 2008), Bird Song (Briggs, Fern, and Raich 2012), and Soccer Player (Zeng et al. 2013). |
| Dataset Splits | Yes | On each dataset, ten-fold cross-validation is performed and the mean accuracy, as well as standard deviation, are reported. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., Python, PyTorch, or specific libraries). |
| Experiment Setup | Yes | In the proposed UCL approach, four trade-off parameters τh, τl, α and ρ need to be manually searched. Figure 2 shows how four parameters affect classification accuracy in four datasets (Lost, MSRCv2, glass, tmc2007)... Specifically, we vary τl ranging from 0.1 to 0.4 when τh = 0.6, α = 0.5, ρ = 0.2; vary τh ranging from 0.5 to 0.9 when τl = 0.2, α = 0.5, ρ = 0.2; vary α ranging from 0.1 to 1.0 when τl = 0.2, τh = 0.6, ρ = 0.2; vary ρ ranging from 0.1 to 0.4 when τl = 0.2, τh = 0.6, α = 0.5. PL-SVM (Nguyen and Caruana 2008): a maximum margin approach that learns from PL examples by optimizing margin-based objective function [suggested configuration: λ 10 3, 10 2 . . . , 103 ]. SURE (Feng and An 2019): a self-training based approach which trains the desired model and performs pseudo-labeling jointly by solving a convex-concave optimization problem [suggested configuration: regularization parameters λ = 0.3, β = 0.05]. |