One-Pass Feature Evolvable Learning with Theoretical Guarantees

Authors: Cun-Yuan Xing, Meng-Zhang Qian, Wu-Yang Chen, Wei Gao, Zhi-Hua Zhou

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We finally conduct extensive experiments to validate the effectiveness of our OPFES method in comparison with the state-of-the-art methods on feature evolvable learning, i.e., our method achieves better performance and the fastest convergence simultaneously. Section 5. Empirical Study We conduct experiments on 20 datasets2, and the details are summarized in Table 1.
Researcher Affiliation Academia 1National Key Laboratory for Novel Software Technology, Nanjing University, China; School of Artificial Intelligence, Nanjing University, China. Correspondence to: Wei Gao <EMAIL>.
Pseudocode Yes Algorithm 1 One-pass optimization of Eqn. (6) Algorithm 2 The OPFES method
Open Source Code Yes 1The code is available at github.com/WeltXing/opfes
Open Datasets Yes We conduct experiments on 20 datasets2, and the details are summarized in Table 1. Most of the datasets have been well-studied for previous feature evolvable learning, and all features have been scaled to [0, 1]. 2Downloaded from Open ML and UCI datasets repository
Dataset Splits Yes For each dataset, we randomly split the feature space into old feature space X [1] and new feature space X [2] with almost equal number of features, following (Gu et al., 2022; Ni et al., 2024). We set Te = 1000 for datasets with a size larger than 10000; otherwise, set Te as 10% of the dataset s size. We also set T1 and T2 as half of the amount of dataset size subtracting Te.
Hardware Specification No No specific hardware details (like GPU/CPU models, memory amounts, or cloud instance types) were mentioned in the paper.
Software Dependencies No The paper mentions techniques like 'online kernel learning with random Fourier features' but does not specify any software libraries or frameworks with version numbers used for implementation.
Experiment Setup Yes For rff-ROGD, rff-FESL, align-FESL and our OPFES, we fix the dimensionality of random Fourier feature as 1000. In the previous stage, we employ Gaussian kernels with widths in 2[−6:6] for all methods. For OPFES, we set TM = 1000 of the optimal stepsize from Theorem 3.7. The stepsize τt is constrained within 10[−4:2]/ t, and the regularization parameter λ is selected from 10[−10:1]. For OCDS, α and β are chosen form 10[−5:0] by cross validations.