Predicting Label Distribution from Multi-label Ranking

Authors: Yunan Lu, Xiuyi Jia

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we implement extensive experiments to validate our proposal.
Researcher Affiliation Academia Yunan Lu, Xiuyi Jia School of Computer Science and Engineering Nanjing University of Science and Technology, Nanjing 210094, China {luyn, jiaxy}@njust.edu.cn
Pseudocode Yes Algorithm 1 Generic DRAM Require: training set {(xn, σn)}N n=1, testing instance x , score function φ, number of mixture components K, number of Monte Carlo samples L;
Open Source Code Yes The supplemental material provides a detailed instruction for reproducing the main results.
Open Datasets Yes We adopt several widely used label distribution datasets, including Movie [4], Emotion6 [18], Twitter LDL, and Flickr-LDL [36].
Dataset Splits Yes Each method is run for ten times on random dataset partitions (70% for training and 30% for test); the average values and standard derivations are recorded. For our method, we set K = 3 and L = 20, and λ is selected from {10 5, 5 10 5, 10 4, 5 10 4, , 101, 5 101} by five-fold cross-validation. For the above comparison methods, since the label distributions are unavailable during training, the hyperparameter configuration that gives the highest Rho on the validation set will be used. On the other hand, we directly train DM and SA on the ground-truth label distributions for comparison. We refer to these two as GT+DM and GT+SA for short, respectively. For these two comparison methods, the hyperparameter configuration that gives the best Cheb, Canber, Cosine, and Rho on the validation set will be used.
Hardware Specification No Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No]
Software Dependencies No The paper mentions software components and algorithms like GL, SA, VI, DM, and L-BFGS, but does not provide specific version numbers for any of them.
Experiment Setup Yes For our method, we set K = 3 and L = 20, and λ is selected from {10 5, 5 10 5, 10 4, 5 10 4, , 101, 5 101} by five-fold cross-validation.