Dynamic Correlation Clustering in Sublinear Update Time

Authors: Vincent Cohen-Addad, Silvio Lattanzi, Andreas Maggiori, Nikos Parotsidis

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally we complement our theoretical analysis with experiments on real world data. and We conduct two sets of experiments.
Researcher Affiliation Collaboration 1Google Research 2Columbia University.
Pseudocode Yes Algorithm 1 AGREEMENTALGORITHM(G), Algorithm 2 Notify(u, ϵ), Algorithm 3 Dynamic Agreement (DA), Algorithm 4 Connect(u, ϵ, t), Algorithm 5 Anchor(u, ϵ, t), Algorithm 6 Clean(u, ϵ, t), Algorithm 7 Compute Connected Components( e G), Algorithm 8 PROBABILISTICAGREEMENT(u, v, ϵ), Algorithm 9 HEAVY(u, ϵ)
Open Source Code Yes Our code is written in Python 3.11.5 and is available at https://github.com/andreasr27/DCC.
Open Datasets Yes We use four graphs from SNAP (Leskovec & Krevl, 2014)... and we use Drift (Vergara et al., 2012; Rodriguez Lujan et al., 2014) from the UCI Machine Learning Repository (Dua & Graff, 2017).
Dataset Splits No The paper describes how dynamic node streams are created ('we first create a random arrival sequence for all the nodes. Subsequently in between any two additions, with probability p, we select at random a node of the current graph and delete it.'), but does not specify train/validation/test splits as the task is dynamic clustering on evolving graphs.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or memory used for running experiments.
Software Dependencies Yes Our code is written in Python 3.11.5
Experiment Setup Yes We set the deletion probability in between any two node arrivals to be 0.2. The agreement parameter is set to ϵ = 0.2... all our subroutines use a random sample of size 2 and the probability of a node joining the anchor set is set to 20/du...