Collaborative Algorithms for Online Personalized Mean Estimation
Authors: Mahsa Asadi, Aurélien Bellet, Odalric-Ambrym Maillard, Marc Tommasi
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We analyze its time complexity and introduce variants that enjoy good performance in numerical experiments. Our numerical experiments on synthetic data are in line with our theoretical findings and show that some empirical variants of our approach can further improve the performance in practice. In this section, we provide numerical experiments on synthetic data to illustrate our theoretical results and assess the practical performance of our proposed algorithms. |
| Researcher Affiliation | Academia | Mahsa Asadi EMAIL Univ. Lille, Inria, CNRS, Centrale Lille UMR 9189 CRISt AL, F-59000 Lille, France; Aurélien Bellet EMAIL Univ. Lille, Inria, CNRS, Centrale Lille UMR 9189 CRISt AL, F-59000 Lille, France; Odalric-Ambrym Maillard EMAIL Univ. Lille, Inria, CNRS, Centrale Lille UMR 9189 CRISt AL, F-59000 Lille, France; Marc Tommasi EMAIL Univ. Lille, Inria, CNRS, Centrale Lille UMR 9189 CRISt AL, F-59000 Lille, France |
| Pseudocode | Yes | Algorithm 1 Col ME Parameters: agent a, time horizon H, risk δ, weighting scheme α, and query strategy choose_agent |
| Open Source Code | Yes | The code can be found at https://github.com/llvllahsa/Collaborative Personalized Mean Estimation |
| Open Datasets | Yes | We consider A = 200 agents, a time horizon of 2500 steps and a risk parameter δ = 0.001. The personal distributions of agents are all Gaussian with variance σ2 = 0.25 and belong to one of 3 classes with means 0.2, 0.4 and 0.8. The class membership of each agent (and thus the value of its true mean) is chosen uniformly at random among the three classes. |
| Dataset Splits | No | The paper uses synthetic data generated according to specified parameters, which is evaluated over '20 random runs corresponding to 20 different samples.' However, it does not provide explicit training/test/validation dataset splits, which are typically used for pre-existing datasets. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models) used for running the experiments. |
| Software Dependencies | No | The paper does not explicitly list specific software dependencies or their version numbers. |
| Experiment Setup | Yes | We consider A = 200 agents, a time horizon of 2500 steps and a risk parameter δ = 0.001. The personal distributions of agents are all Gaussian with variance σ2 = 0.25 and belong to one of 3 classes with means 0.2, 0.4 and 0.8. The class membership of each agent (and thus the value of its true mean) is chosen uniformly at random among the three classes. All algorithms are compared across 20 random runs corresponding to 20 different samples. |