Max-Margin Majority Voting for Learning from Crowds

Authors: TIAN TIAN, Jun Zhu

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now present experimental results to demonstrate the strong discriminative ability of max-margin majority voting and the promise of our Bayesian models, by comparing with various strong competitors on multiple real datasets.
Researcher Affiliation Academia Tian Tian, Jun Zhu Department of Computer Science & Technology; Center for Bio-Inspired Computing Research Tsinghua National Lab for Information Science & Technology State Key Lab of Intelligent Technology & Systems; Tsinghua University, Beijing 100084, China tiant13@mails.tsinghua.edu.cn; dcszj@tsinghua.edu.cn
Pseudocode Yes Algorithm 1: The Crowd SVM algorithm
Open Source Code No The paper does not contain an unambiguous statement or link indicating that the source code for the methodology described in this paper is publicly available.
Open Datasets Yes We use four real world crowd labeling datasets as summarized in Table 1. Web Search [24]: ... Age [8]: ... Bluebirds [19]: ... Flowers [18]:
Dataset Splits No The paper states: 'we cannot simply split the training data into multiple folds to cross-validate the hyperparameters' and instead uses likelihood on the given dataset for model selection, without specifying any train/validation/test splits.
Hardware Specification Yes All experiments were conducted on a PC with Intel Core i5 3.00GHz CPU and 12.00GB RAM.
Software Dependencies No The paper mentions using 'well-developed SVM solvers like LIBSVM [2]' but does not provide specific version numbers for LIBSVM or any other software dependencies.
Experiment Setup Yes For Crowd SVM, we set α = 1 and v = 1 for all experiments, since we find that the results are insensitive to them. For M3V, Crowd SVM and Gibbs-Crowd SVM, the regularization parameters (c, ℓ) are selected from c = 2ˆ[ 8 : 0] and ℓ= [1, 3, 5] by the method in Sec. 6.2. As for Gibbs-Crowd SVM, we generate 50 samples in each run and discard the first 10 samples as burn-in steps