Poster M26: Spectral Clustering with Perturbed Data
Ling Huang, Donghui Yan, Michael I. Jordan and Nina Taft

Motivation: To provide a general analytic framework to analyze the impact of data perturbation on spectral clustering. Perturbing data via filtering, quantization, or synopsis-based approximation is popular for communication and computation efficiencies. Our Solution In-Network Efficient Learning Framework End devices inputs: error Bounds end-to-end tolerance error propagation. data1(t) reduced_data1(t) Quantifies tradeoffs Reasoning between data perturbation Center perturbation and data2(t) analysis clustering accuracy. reduced_data2(t) Provides accuracy learning guarantees while inference enabling data datan(t) adjust sys. reduction in practical parameters settings.
reduced_datan(t) 1