Customizing ML Predictions for Online Algorithms

Authors: Keerti Anand, Rong Ge, Debmalya Panigrahi

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

Reproducibility Variable Result LLM Response
Research Type Experimental We support this fnding both through the oretical bounds and numerical simulations. We use numerical simulations to evaluate the algorithms that we designed for the learning-to-rent prob lem: the black box algorithm (Algorithm 3), the marginbased algorithm (Algorithm 4), and the algorithm for a noisy classifer (Algorithm 5).
Researcher Affiliation Academia Department of Computer Science, Duke University, Durham NC, United States.
Pseudocode Yes Algorithm 1 Outputs θA for a given distribution on y. Algorithm 2 Outputs θA(x) for multi-dimensional x. Algorithm 3 Black box learning-to-rent algorithm. Algorithm 4 Margin-based learning-to-rent algorithm. Algorithm 5 Learning-to-rent with a noisy classifer.
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository.
Open Datasets No Experimental Setup. We frst describe the joint distribu tion (x, y) K used in the experiments. We choose a random vector W Rd as W N(0, I/d). We view W as a hyper-plane passing through the origin (W T x = 0). ...The input x is drawn from a mixture distribution, where with probability 1/2 we sample x from a Gaussian x N(0, I/d), and with probability 1/2, we sample x as x = αW + η, here α N(0, 1) is a coeffcient in the direction 1 of W and η N(0, I). Choosing x from the Gaussian d distribution ensures that the data-set has no margin; however, in high dimensions, W T x will concentrate in a small region, which makes all the label y very close to 1. We address this issue by mixing in the second component which ensures that the distribution of y is diverse. No direct access (link, DOI, or specific citation for public availability) is provided for the custom-generated dataset used in the experiments.
Dataset Splits Yes Training and Validation. For a given training set, we split it in two equal halves, the frst half is used to train our PAC learner and the second half is used as a validation set to optimize the design parameters in the algorithms, namely τ in Algorithm 3 and γ in Algorithm 4.
Hardware Specification No The paper does not specify the exact hardware (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific libraries).
Experiment Setup Yes Experimental Setup. We frst describe the joint distribu tion (x, y) K used in the experiments. We choose a random vector W Rd as W N(0, I/d). ... The input x is drawn from a mixture distribution, where with probability 1/2 we sample x from a Gaussian x N(0, I/d), and with probability 1/2, we sample x as x = αW + η, here α N(0, 1) is a coeffcient in the direction 1 of W and η N(0, I). ...For a given training set, we split it in two equal halves, the frst half is used to train our PAC learner and the second half is used as a validation set to optimize the design parameters in the algorithms, namely τ in Algorithm 3 and γ in Algorithm 4.