Multi-slots Online Matching with High Entropy

Authors: Xingyu Lu, Qintong Wu, Wenliang Zhong

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on synthetic and industrial data sets demonstrate that OG-MA is a fast and promising method for multi-slots online matching. Extensive experiments on synthetic and industrial data sets validate our computation complexity analysis and sub-linear regret results. Further, we show that OG-MA algorithm achieves positive diversity uplift by the trade-off between the entropy regularizer and the matching revenue.
Researcher Affiliation Industry 1Ant Group, Hangzhou, China. Correspondence to: Xingyu Lu <sing.lxy@antgroup.com>, Qintong Wu <qintong.wqt@antgroup.com>, Wenliang Zhong <yice.zwl@antgroup.com>.
Pseudocode Yes Algorithm 1 Online sub Gradient descent for Multi-slots Allocation (OG-MA), Algorithm 2 Efficiency Pooling Projection (EPP), Algorithm 3 Roulette Swapping Allocation (RSA)
Open Source Code No The paper does not include an unambiguous statement that the authors are releasing the code for the work described in this paper, nor does it provide a direct link to a source-code repository.
Open Datasets No The paper mentions 'Industrial Dataset. We utilize Ali Express Searching System Dataset France to evaluate the performance of the proposed algorithm on a large-scale problem.' While the dataset name is mentioned, no link, DOI, repository name, or formal citation with author names and year in brackets/parentheses is provided to indicate public availability.
Dataset Splits No The paper describes generating a user set of T=10000 by random sampling and permuting it, but it does not specify explicit train, validation, or test dataset splits with percentages, absolute sample counts, or references to predefined splits.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, processor types, or detailed computer specifications used for running its experiments. It generally refers to 'an online server'.
Software Dependencies No The paper mentions using 'CVX package (Diamond & Boyd, 2016) with MOSEK solver (Ap S, 2022)' for implementing a baseline (DSDA), but it does not provide specific version numbers for ancillary software dependencies used for their proposed OG-MA algorithm.
Experiment Setup Yes We choose the regularization level α = 0.01 and decay factor γ = 1/2. We plot the mean regret with 95% confidence interval over 100 random trials for each value of T and set α = 0.01. We choose the regularization level α {10-5, 10-4, 10-3, 10-2, 10-1, 1} in the experiment. We employed a two-layers Multi-Layer Perceptron (MLP) to parameterize the revenue function f(t, a). Specifically, we set embedding size demb = 8 for each sparse feature, learning rate = 0.001, batch size = 256, l2 regularizer = 0.001, and select Adam as the optimizer. The two-layers MLP is of [32, 16] hidden units.