search

squareRoot3 / Rethinking-Anomaly-Detection

"Rethinking Graph Neural Networks for Anomaly Detection" in ICML 2022
https://proceedings.mlr.press/v162/tang22b.html
172 stars 27 forks source link

Why band-pass filter can alleivate the "right shift" effect due to graph anomalies in the first place? #11

Closed jingweio closed 8 months ago

jingweio commented 9 months ago

Hi, I would like to first express my appreciation for the theoretical analysis presented in the first half of your paper. Your work from a spectral perspective provides a insightful quantification of the impact of anomalies on a clean graph.

However, moving on to the latter part of the paper, I find myself somewhat uncertain about a few aspects. While I understand how beta kernels shift from left to right to capture high-frequency information, the mechanics of how a band-pass filter counters the "right shift" effect due to graph anomalies remains unclear to me.

Why did you choose a band-pass filter as your preferred option? Is it because it can effectively capture the nuances between low and high-frequency information by easily shifting towards either the left or right sides? Could you kindly elaborate on the motivation and thought process behind your model's design in this regard?

squareRoot3 commented 8 months ago

Is it because it can effectively capture the nuances between low and high-frequency information by easily shifting towards either the left or right sides?

Yes.