Hello，it's really confusing that GP performed very very well in modeling this data,but DeepGP didn't get such a good result.I'm looking forward to your reply，thanks！

[pingpingLv]

[bydeng01]

It really depends on the nature and size of dataset. GP as a non-parametric(infinite latent variables) method, directly models all the data points without approximations, making it very flexible and effective for smaller or moderately sized datasets.

While, DGP usually rely on inducing points and sparse variational inference to approximate the posterior. This approximation allows computational efficiency for large dataset by using a finite set of inducing points rather than directly modeling all points. So, it can sometimes cause DGP to underfit small or sparse datasets, leads to less accurate results compared to a standard GP.

I would recommend you to read Doubly Stochastic Variational Inference for Deep Gaussian Processes which explain from GP to Deep GP pretty well