Self-Paced Unified Representation Learning for Hierarchical Multi-Label Classification
Authors: Zixuan Yuan, Hao Liu, Haoyi Zhou, Denghui Zhang, Xiao Zhang, Hao Wang, Hui Xiong
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on several empirical benchmarks demonstrate the effectiveness and efficiency of SPUR compared to state-of-the-art methods, especially in scenarios with missing features. |
| Researcher Affiliation | Collaboration | Zixuan Yuan1, Hao Liu1, Haoyi Zhou2, Denghui Zhang3, Xiao Zhang4, Hao Wang5, Hui Xiong1 1Hong Kong University of Science and Technology (Guangzhou) 2Beihang University, 3Stevens Institute of Technology, 4Shandong Univerisity, 5Alibaba Group |
| Pseudocode | Yes | Algorithm 1: Training procedure for SPUR framework. 1: Input: The training set {X, Y}, structure-aware learner F( ; θ), learning rate η, update frequency K of difficulty measurement 2: Output: the well-trained parameters θ of structureaware learner 3: while not converge do 4: epoch = 1 5: for (x, y) {X, Y} do 6: initialize Lgd = 7: for l = 1, . . . , L do 8: Obtain the training loss L(l) based on the Eq. (5) 9: Lgd.add(L(l)) 10: end for 11: if epoch! = 1 then 12: Compute the likelihood wx of Lgd being assigned to the majority category 13: 14: θ θ η(P x X wx Ltg x λ P x X wx) based on Eq. (7) 15: end if 16: end for 17: if epoch == 1 or epoch%K == 0 then 18: Update ˆL(l), l {1, . . . , L} 19: end if 20: epoch = epoch + 1 21: end while |
| Open Source Code | Yes | Code released at https://github.com/yuanzx33033/SPUR/ |
| Open Datasets | Yes | We evaluate the generic efficacy of our SPUR on 20 publicly-available datasets, which include protein function prediction (Ruepp et al. 2004; Ashburner et al. 2000), annotation of medical (Dimitrovski et al. 2011) or microalgae images (Dimitrovski et al. 2012), and text categorization (Klimt and Yang 2004). These datasets are commonly used for comparison among HMLC approaches (Giunchiglia and Lukasiewicz 2020; Bi and Kwok 2011; Nakano, Lietaert, and Vens 2019; Wehrmann, Cerri, and Barros 2018). |
| Dataset Splits | No | The paper mentions a 'training set' in Algorithm 1 and lists 20 datasets, but it does not explicitly provide specific training/validation/test dataset split percentages, absolute sample counts for each split, or detailed splitting methodology. |
| Hardware Specification | Yes | Our SPUR1 and all baselines are implemented with Pytorch framework (Paszke et al. 2019) and run on a single 3090 Ti GPU. |
| Software Dependencies | No | The paper states 'implemented with Pytorch framework (Paszke et al. 2019)' but does not provide a specific version number for Pytorch or any other software dependencies. |
| Experiment Setup | No | The paper mentions general training aspects like 'learning rate η, update frequency K of difficulty measurement' in Algorithm 1, and 'the number of learning groups to be two', but it does not provide specific hyperparameter values (e.g., actual values for η, K, batch size, number of epochs, optimizer settings) or explicit system-level training configurations in the main text. |