Constrained Interacting Submodular Groupings

Authors: Andrew Cotter, Mahdi Milani Fard, Seungil You, Maya Gupta, Jeff Bilmes

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We close with a case study in which we use these algorithms to find high quality diverse ensembles of classifiers, showing good results. We validate our proposed approach with a case study in the setting of Canini et al. (2016). We compare to two baselines, the Crystals and Random Tiny Lattice (RTL) algorithms of Canini et al. (2016).
Researcher Affiliation Collaboration 1Google AI 2Kakao Mobility 3University of Washington, Seattle, work done while at Google AI.
Pseudocode Yes Algorithm 1 Adaptation of the General Greed SAT algorithm of Wei et al. (2015) to handle matroid constraints, crossblock interactions, and coverings/packings as well as partitions.
Open Source Code No The paper does not provide any explicit statements or links for open-source code for the methodology it describes.
Open Datasets No The dataset contains 463 154 samples with 29 informative features plus a binary label indicating whether a particular visual element should be displayed on a web page. The paper does not provide specific access information (link, DOI, or citation to a public repository) for this dataset.
Dataset Splits Yes The dataset was randomly partitioned into training, validation and testing subsets containing 80%, 10% and 10% of the data, respectively (the validation set was only used for hyperparameter optimization of the baseline Crystals algorithm).
Hardware Specification Yes Each optimization took between 2 and 30 seconds on a Xeon E5-2690.
Software Dependencies No The paper mentions algorithms and models used (e.g., 'randomized algorithm of Feldman et al. (2017)'), but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, or specific library versions).
Experiment Setup Yes Our goal here is essentially identical to Canini et al. (2016) we seek to choose m = 8 lattices, each containing up to 8 features (via a matroid constraint)... We optimized Eq. (6) for various choices of λ4 using the randomized algorithm of Feldman et al. (2017) combined with the procedure of Lemma 4, with β = 0.5 and δ = 0.1.