Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Collective Intelligence in Decision-Making with Non-Stationary Experts

Authors: Axel Abels, Vito Trianni, Ann Nowé, Tom Lenaerts

JAIR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct an extensive empirical evaluation of our novel method and compare it to a range of baselines in Section 6.1. Through this evaluation we demonstrate that our proposed method provides a significant improvement in performance over previous adaptive algorithms for a wide variety of configurations. We further show that unlike previous algorithms, which require their adaptiveness to be tuned to match changes in expertise, our novel approach is more robust in terms of its adaptiveness parameter. We conclude by applying our methods to active learning in Section 6.2, demonstrating improved performance on a concrete real-world problem.
Researcher Affiliation Academia AXEL ABELS , Machine Learning Group, Université Libre de Bruxelles, Belgium, AI Lab, Vrije Universiteit Brussel, Belgium, and FARI Institute, Université Libre de Bruxelles Vrije Universiteit Brussel, Belgium VITO TRIANNI, Institute of Cognitive Sciences and Technologies, National Research Council, Italy ANN NOWÉ, AI Lab, Vrije Universiteit Brussel, Belgium and FARI Institute, Université Libre de Bruxelles Vrije Universiteit Brussel, Belgium TOM LENAERTS, Machine Learning Group, Université Libre de Bruxelles, Belgium, AI Lab, Vrije Universiteit Brussel, Belgium, Center for Human-Compatible AI, UC Berkeley, USA, and FARI Institute, Université Libre de Bruxelles Vrije Universiteit Brussel, Belgium
Pseudocode Yes Algorithm 1 CORVAL
Open Source Code Yes Code to reproduce these results is available at https://github.com/axelabels/CDM_NONSTAT.
Open Datasets Yes Following previous works [30, 12, 23, 22], we evaluate performance on 20 data sets which have previously been used to evaluate active learning approaches, namely bank-marketing, calhousing, cod-rna, credit-g, diabetes, eeg-eye-state, electricity, ibn-sina, ijcnn1, kc2, kdd99_10perc, magic Telescope, mozilla4, musk, ozone-level-8hr, qsar-biodeg, steel-plates-fault, svmguide3, tic-tac-toe, and zebra.
Dataset Splits Yes For each data set, we set aside 1/3𝑟𝑑of the data points as test set and run the active learning set-up for 100 steps on the remaining data.
Hardware Specification No The resources and services used in this work were provided by the VSC (Flemish Supercomputer Center), funded by the Research Foundation Flanders (FWO) and the Flemish Government.
Software Dependencies No Following Chu and Lin 2016, we train a Logistic Regression classifier [21].
Experiment Setup Yes We evaluate the performance of the chosen algorithms in terms of reward over 𝑇= 5000 steps averaged over 200 simulations4. For a given period (𝜏 {100, 500, 2500}), we generate non-stationary expertise by averaging 100 randomly sampled sine waves each with period ˆ𝜏 N (𝜏,𝜏/2). For each data set, we set aside 1/3𝑟𝑑of the data points as test set and run the active learning set-up for 100 steps on the remaining data.