Conditional Bernoulli Mixtures for Multi-label Classification

Authors: Cheng Li, Bingyu Wang, Virgil Pavlu, Javed Aslam

ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show the effectiveness of the proposed method against competitive alternatives on benchmark datasets. We perform experiments on five commonly used and relatively large multi-label datasets: SCENE, TMC2007, MEDIAMILL, NUS-WIDE from Mulan1 and RCV1 (topics subset 1) from LIBSVM2.
Researcher Affiliation Academia Cheng Li CHENGLI@CCS.NEU.EDU Bingyu Wang RAINICY@CCS.NEU.EDU Virgil Pavlu VIP@CCS.NEU.EDU Javed Aslam JAA@CCS.NEU.EDU College of Computer and Information Science, Northeastern University, Boston, MA 02115, USA
Pseudocode Yes Algorithm 1 Generic Training for CBM and Algorithm 2 Prediction by Dynamic Prog. and Pruning
Open Source Code Yes Our implementations of CBM and several baselines (Pow Set, PCC, CRF, etc.) are available at https://github.com/ cheng-li/pyramid.
Open Datasets Yes We perform experiments on five commonly used and relatively large multi-label datasets: SCENE, TMC2007, MEDIAMILL, NUS-WIDE from Mulan1 and RCV1 (topics subset 1) from LIBSVM2. 1http://mulan.sourceforge.net 2https://www.csie.ntu.edu.tw/~cjlin/ libsvmtools/datasets/multilabel.html
Dataset Splits Yes For the sake of reproducibility, we adopt the train/test splits provided by Mulan and LIBSVM. Hyper parameter tuning is done by cross-validation on the training set (see the supplementary material for details).
Hardware Specification No The paper does not provide any specific hardware details such as GPU/CPU models, memory, or cloud instance types used for the experiments.
Software Dependencies No The paper mentions software like logistic regressions, gradient boosted trees, L-BFGS, MEKA, LIBSVM, and Python, but does not specify any version numbers for these or other software dependencies.
Experiment Setup Yes To avoid over-fitting, we also add L2 regularizations (Gaussian priors) to all parameters. Hyper parameter tuning is done by cross-validation on the training set (see the supplementary material for details).