Multi-agent Performative Prediction with Greedy Deployment and Consensus Seeking Agents

Authors: Qiang LI, Chung-Yiu Yau, Hoi-To Wai

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical results validate our analysis. We consider two examples of performative prediction problems to verify our theories. All experiments are conducted with Python on a server using 80 threads of an Intel Xeon 6318 CPU. Multi-agent Gaussian Mean Estimation. We aim to illustrate Proposition 1, Theorem 1 via a scalar Gaussian mean estimation problem on synthetic data. Email Spam Classification. We evaluate the performance of DSGD-GD by simulating the performative effects on a real dataset.
Researcher Affiliation Academia Qiang Li Chung-Yiu Yau Hoi-To Wai Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong, Shatin, Hong Kong SAR of China {liqiang, cyyau, htwai}@se.cuhk.edu.hk
Pseudocode Yes DSGD with Greedy Deployment (DSGD-GD) Scheme At iteration t = 0, 1, ..., for any i V , agent i updates his/her decision (θt i) by the recursion consisting of two phases (Phase 1) Zt+1 i Di(θt i) (Phase 2) θt+1 i = Pn j=1 Wijθt j γt+1 ℓ(θt i; Zt+1 i ), (5)
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the described methodology is open-source or publicly available.
Open Datasets Yes Email Spam Classification. This example is a multi-agent spam classification task based on spambase, a dataset [Hopkins, 1999] with m = 4601 samples, d = 48 features.
Dataset Splits Yes Each server has access to training data from mi = 138 samples from spambase modeling the different set of users; the rest of mtrain = 1150 samples are taken as testing data.
Hardware Specification Yes All experiments are conducted with Python on a server using 80 threads of an Intel Xeon 6318 CPU.
Software Dependencies No The paper mentions
Experiment Setup Yes In our experiments, we set zi = 10, σ2 = 50 and step size for DSGD-GD as γt = a0/(a1 + t) with a0 = 50, a1 = 104. The sensitivity parameters are set as ϵi {0.4ϵavg, 0.45ϵavg, ..., 1.6ϵavg} with ϵ = ϵavg {0.01, 0.1, 1}. The servers aim to find a common spam filter classifier via (2) with β = 10−4.