Enhancing Semi-supervised Domain Adaptation via Effective Target Labeling

Authors: Jiujun He, Bin Liu, Guosheng Yin

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct thorough evaluations on three image-based benchmark datasets: Office-31 (Saenko et al. 2010), Office-Home (Venkateswara et al. 2017), and Domain Net (Peng et al. 2019). The results are reported in Tables 2 and 3.
Researcher Affiliation Academia 1Center of Statistical Research, School of Statistics, Southwestern University of Finance and Economics, Chengdu, China 2Department of Statistics and Actuarial Science, The University of Hong Kong, Hong Kong, China
Pseudocode Yes Algorithm 1: Overview of our proposed learning framework and Algorithm 2: Non-maximal degree node suppression
Open Source Code Yes For more details, please refer to our code: https://github.com/BetterTMrR/EFTL-Pytorch-main.
Open Datasets Yes We conduct thorough evaluations on three image-based benchmark datasets: Office-31 (Saenko et al. 2010), Office-Home (Venkateswara et al. 2017), and Domain Net (Peng et al. 2019).
Dataset Splits No No explicit train/validation/test dataset splits (percentages or counts) or references to standard predefined splits are mentioned for the overall datasets used (Office-31, Office-Home, Domain Net).
Hardware Specification No No specific hardware details (GPU/CPU models, processor types, memory amounts) are mentioned for the experimental setup.
Software Dependencies No No specific software dependencies with version numbers are provided. The paper mentions 'Pytorch' in the GitHub link but without a version.
Experiment Setup Yes Since the numbers of the AEN and ADN in the directed graph construction depend on the scale of the dataset, we choose M1 = nt/(3 K) and M2 = M1/5 to adaptively construct the directed graph across all datasets. The trade-off hyperparameter α in Eq. (4) is fixed as 0.1. For baseline Fix MME, the threshold τ in Eq. (7) is set to be 0.85 for all datasets except 0.8 for Domain Net. Following Li et al. (2021b), we exploit a label smoothing technique with parameter 0.1 to avoid overconfident predictions when using a cross-entropy loss. We run our experiments three times with different random seeds independently.