Gradient-Based Optimization for Bayesian Preference Elicitation

Authors: Ivan Vendrov, Tyler Lu, Qingqing Huang, Craig Boutilier10292-10301

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our gradientbased EVOI technique achieves state-of-the-art performance across several domains while scaling to large item spaces.
Researcher Affiliation Industry Ivan Vendrov, Tyler Lu, Qingqing Huang, Craig Boutilier Google Research, Mountain View, California {ivendrov, tylerlu, qqhuang, cboutilier}@google.com
Pseudocode Yes Algorithm 1 Deep Retr Uniq. Inputs: optimized X and U
Open Source Code No The paper mentions using "Tensor Flow, Py Torch" as computational frameworks but does not provide any statement about releasing their own source code or a link to a repository.
Open Datasets Yes Using the Movie Lens 100-K dataset, we train user and movie embeddings with dimension d = 10. ... We use the Movie Lens-20M (Harper and Konstan 2015) dataset and represent each movie with 100 binary attributes from the Tag Genome (Vig, Sen, and Riedl 2012).
Dataset Splits No The paper uses datasets like Movie Lens 100-K and Goodreads, and describes running elicitation trials with random selections of items and user embeddings, but it does not provide specific train/validation/test dataset split percentages or counts.
Hardware Specification Yes We benchmark the algorithmic runtimes on a workstation with a 12-core Intel Xeon E5-1650 CPU at 3.6GHz, and 64GB of RAM.
Software Dependencies No The paper mentions using "Tensor Flow, Py Torch" as computational frameworks but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes We initialize query embeddings to random uniform values in [0, 1]100, then run gradient ascent on Eq. 7 for 100 steps, initializing the regularization weight λ at 0.01 and multiplying λ by 1.1 each iteration.