Manifold Regularization
Removing manifold regularization reduced alive features and lowered conversion-prediction AUC. The manifold penalty is applied to pre-activations so that even features removed by TopK can receive gradients. GeoSAE builds a k-nearest-neighb…
1 sources - 4 claims
Removing manifold regularization reduced alive features and lowered conversion-prediction AUC. The manifold penalty is applied to pre-activations so that even features removed by TopK can receive gradients. GeoSAE builds a k-nearest-neighbor graph over layer representations using Gaussian kernel weights. Too much manifold regularization over-regularized the representation and degraded performance.