The　anisotropic　kernel　ridge　regression　(AKRR)　approach　in　nuclear　mass　predictions　is　developed　by　introducing　the　anisotropic　kernel　function　into　the　kernel　ridge　regression　(KRR)　approach,　without　introducing　new　weight　parameter　or　input　in　the　training.　A　combination　of　double　two-dimensional　Gaussian　kernel　function　is　adopted,　and　the　corresponding　hyperparameters　are　tuned　carefully　through　cross-validations　to　optimize　the　predictions.　The　anisotropic　kernel　shows　cross-shape　pattern,　which　highlights　the　correlations　among　the　isotopes　with　the　same　proton　number,　and　those　among　the　isotones　with　the　same　neutron　number.　Significant　improvements　are　achieved　by　the　AKRR　approach　in　both　the　interpolation　and　the　extrapolation　predictions　of　nuclear　masses　comparing　with　the　original　isotropic　KRR　approach.　©　2024　American　Physical　Society.