Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Uncertain Knowledge Graph Completion via Semi-Supervised Confidence Distribution Learning
Authors: Tianxing Wu, Shutong Zhu, Jingting Wang, Ning Xu, Guilin Qi, Haofen Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present experiments to show the effectiveness and superiority of ss CDL on the UKG completion tasks of confidence prediction and link prediction. We also analyze the effects of CDL and meta self-training in ss CDL with ablation experiments, and further explore the performance of ss CDL on predicting the confidences of low-confidence triples. |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, Southeast University, China 2Key Laboratory of New Generation Artificial Intelligence Technology and its Interdisciplinary Applications (Southeast University), Ministry of Education, China 3College of Design and Innovation, Tongji University, China EMAIL |
| Pseudocode | Yes | Algorithm 1 gives the details on meta self-training of ss CDL. |
| Open Source Code | Yes | The source code of ss CDL is publicly available at: https://github.com/seucoin/un KR/tree/main/un KR_ss CDL. |
| Open Datasets | Yes | We conducted experiments on two widely used UKG datasets, i.e., NL27k extracted from NELL and CN15k extracted from Concept Net (more details are in Table 1). |
| Dataset Splits | Yes | We followed the setting of UKGE [3] to partition the dataset into 85% for training, 7% for validation, and 8% for testing. |
| Hardware Specification | Yes | ss CDL was implemented by Pytorch-lightning, and all the experiments were conducted on an RTX3090 GPU card. |
| Software Dependencies | No | ss CDL was implemented by Pytorch-lightning, and all the experiments were conducted on an RTX3090 GPU card. We used Adam optimizer [13] for SGD training. |
| Experiment Setup | Yes | The optimal hyper-parameters of ss CDL on NL27k are as follows: the standard deviation of Gaussian distribution σ = 0.6, the weight of pseudo labeled data wp = 0.7, the MSE weight β = 1, the learning rate α = 0.001, the margin γ = 0.1, the threshold for pseudo labeled data selection: 0.03, the batch size: 4096, and the embedding size: 128. The optimal hyper-parameters of ss CDL on CN15k are as follows: the standard deviation of Gaussian distribution σ = 0.6, the weight of pseudo labeled data wp = 0.3, the MSE weight β = 1, the learning rate α = 0.001, the margin γ = 0.1, the threshold for pseudo labeled data selection: 0.015, the batch size: 4096, and the embedding size: 512. |