Optimal Continuous DR-Submodular Maximization and Applications to Provable Mean Field Inference

Authors: Yatao Bian, Joachim Buhmann, Andreas Krause

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the superior performance of our algorithms with baseline results on both synthetic and real-world datasets. Extensive experimental evaluations on real-world and synthetic data support our theory. We tested on the representative FLID model the following algorithms and baselines: The first category is one-epoch algorithms, including 1 Submodular-Double Greedy (Bian et al., 2017c) with 1/3 approximation guarantee, 2 BSCB (Alg. 4 of Niazadeh et al. (2018), where we chose = 10 3) with 1/2 guarantee and 3 DR-Double Greedy (Alg. 1) with 1/2 guarantee.
Researcher Affiliation Academia 1Department of Computer Science, ETH Zurich, Zurich, Switzerland. Correspondence to: Yatao A. Bian <ybian@inf.ethz.ch>.
Pseudocode Yes Algorithm 1: DR-Double Greedy(f, a, b); Algorithm 2: DG-Mean Field-1/2 & DG-Mean Field-1/3
Open Source Code No All algorithms are implemented in Python3, and source code will be public on the author s homepage.
Open Datasets Yes We tested the mean field methods on the trained FLID models from Tschiatschek et al. (2016) on Amazon Baby Registries dataset. After preprocessing, this dataset has 13 categories, e.g., feeding & furniture.
Dataset Splits Yes Split the training data into multiple folds, train a model on each fold D and infer a noisy posterior distribution p(S|D). For each category, three classes of models were trained, with latent dimensions D = 2, 3, 10, repectively, on 10 folds of the data.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory) used for running the experiments.
Software Dependencies No All algorithms are implemented in Python3, and source code will be public on the author s homepage. The paper mentions Python3 but does not specify version numbers for Python or any other libraries or software dependencies.
Experiment Setup Yes The objectives under investigation are ELBO (1) and PA-ELBO (2). We set β = 1 in PA-ELBO. For all algorithms, we use the same random order to process the coordinates within each epoch. For each category, three classes of models were trained, with latent dimensions D = 2, 3, 10, repectively, on 10 folds of the data.