Marginalized Denoising for Link Prediction and Multi-Label Learning
Authors: Zheng Chen, Minmin Chen, Kilian Weinberger, Weixiong Zhang
AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluated MLLP on three real-world applications one social network and two PPI networks. MLLP outperforms every competing prior work consistently across all our benchmark data sets. |
| Researcher Affiliation | Collaboration | Zheng Chen1,2, Minmin Chen3, Kilian Q. Weinberger2, Weixiong Zhang2,1 1Institute for Systems Biology, Jianghan University, Wuhan, Hubei 430056, China. 2Department of Computer Science and Engineering, Washington University in St. Louis, St. Louis, MO 63130, USA. 3Criteo Lab, Palo Alto, CA, USA |
| Pseudocode | Yes | Algorithm 1 Joint link prediction and multi-label learning |
| Open Source Code | Yes | For perfect reproducibility, we provide the source code and data sets online at http://www.cse.wustl.edu/ chenz/codes/mllp.tar.gz. |
| Open Datasets | Yes | Yeast: we extracted the yeast PPI graph from HINT (Das and Yu 2012)... The functional annotations (labels) of proteins were obtained from the SGD database1. 1http://www.yeastgenome.org/ Human: the human PPI graph was also taken from HINT. The human proteins were annotated according to the Gene Ontology (Consortium and others 2000). Live Journal: ... (Yang and Leskovec 2012). |
| Dataset Splits | Yes | To establish ground truth for comparing performance of the method considered, we removed labels and edges in three ways. 1) We randomly removed a fraction, γ {0.0, 0.1, 0.2, 0.3, 0.4}, of the existing edges... 3) To simulate various degrees of incomplete training information, we randomly removed a fraction δ {0.0, 0.1, 0.2, 0.3, 0.4} of the label information Y of all remaining nodes... We set the hyper-parameters of all methods using 5-fold cross-validation... |
| Hardware Specification | Yes | On a server with two 8-core Intel Xeon E5-2650@2.60GHz CPUs and 128GB RAM, our MATLABT M implementation (accelerated with an NVIDIA K40 GPU) |
| Software Dependencies | No | The paper mentions 'our MATLABTM implementation' but does not specify a version number for MATLAB or any other software libraries used. |
| Experiment Setup | Yes | We removed labels and edges in three ways. 1) We randomly removed a fraction, γ {0.0, 0.1, 0.2, 0.3, 0.4}, of the existing edges. 2) We randomly selected 50% of the nodes and remove all label information from them (these are the MLL test nodes). 3) To simulate various degrees of incomplete training information, we randomly removed a fraction δ {0.0, 0.1, 0.2, 0.3, 0.4} of the label information Y of all remaining nodes... We set the hyper-parameters of all methods using 5-fold cross-validation, except the corruption noise p of MLLP and MDM is set to a high value p=0.95, throughout. |