A General Online Algorithm for Optimizing Complex Performance Metrics

Authors: Wojciech Kotlowski, Marek Wydmuch, Erik Schultheis, Rohit Babbar, Krzysztof Dembczynski

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also verify the efficiency of the proposed algorithm in empirical studies, where we evaluate it on a wide range of multi-label benchmarks and performance measures.
Researcher Affiliation Collaboration 1Poznan University of Technology, Poznan, Poland 2Aalto University, Helsinki, Finland 3University of Bath, Bath, UK 4Yahoo Research, New York, United States.
Pseudocode Yes Algorithm 1 Online Measure Maximization Algorithm; Algorithm 2 OMMA(bη)
Open Source Code Yes Code to reproduce the experiments: https://github.com/mwydmuch/x COLUMNs
Open Datasets Yes For multi-class experiments, we use News20 (Lang, 1995), Ledgar-Lex Glue (Chalkidis et al., 2022) with tf-idf features, Caltech-256 (Griffin et al., 2007) with features obtained using VGG16 (Simonyan & Zisserman, 2014) trained on Image Net, and for multi-label experiments, we use You Tube, Flickr (Tang & Liu, 2009) with Deep Walk features (Perozzi et al., 2014), Eurlex-Lex Glue (Chalkidis et al., 2021), Mediamill (Snoek et al., 2006), RCV1X (Lewis et al., 2004), and Amazon Cat (Mc Auley & Leskovec, 2013; Bhatia et al., 2016) with tf-idf features.
Dataset Splits Yes For benchmarks without default train and test sets, we split them randomly in proportion 70/30.
Hardware Specification No The paper states: 'All the experiments were conducted on a workstation with 64 GB of RAM.' This is a general description and does not provide specific details such as CPU/GPU models, processor types, or exact memory amounts (e.g., type of RAM).
Software Dependencies Yes In multi-label experiments with fixed CPE, we use LIBLINEAR model with L2-regularized logistic loss (Fan et al., 2008). In multi-class experiments with both fixed and online CPE, we use Py Torch (Paszke et al., 2019) linear model with L2-regularized binary cross entropy or cross entropy loss, optimized using ADAM (Kingma & Ba, 2015).
Experiment Setup Yes As discussed in Remark 4.4, we add both small constant = 1e-9 to the denominators in all metrics, as well as the regularization value {0, 1e-6, 1e-3, 0.1, 1} to the entries of the confusion matrix used in all online algorithms (Greedy, OFO, Online-FW, and OMMA).