Popularity　bias　commonly　exists　in　recommendation　systems.　Previous　studies　mainly　evaluated　the　performance　of　recommendation　models　offline　through　normal　bias　testing　of　factual　interactions　or　debiased　testing　using　records　from　randomized　controlled　trials.　Yet,　achieving　high　performance　on　one　of　the　tests　only　provides　a　partial　view.　For　example,　the　debiased　test　eliminates　system-induced　biases,　enabling　the　model　to　strive　for　an　equitable　distribution　of　items　in　recommendation　lists.　However,　this　pursuit　carries　the　risk　of　over-debiasing,　potentially　undermining　recommendation　performance.　In　this　paper,　our　goal　is　to　craft　a　recommendation　method　that　achieves　unbiasedness　and　excels　in　both　tests.　This　necessitates　the　development　of　a　model　capable　of　making　accurate　predictions　in　both　the　debiased　test　and　the　biased　test.　To　this　end,　we　propose　a　novel　Empirical　Risk　Minimization　for　Popularity　Debiasing　(ERMPD)　training　method　for　training　a　test-agnostic　model　to　achieve　unbiasedness,　which　has　truly　excellent　recommendation　performance　while　eliminating　popularity　bias.　From　a　causal　perspective,　ERMPD　uses　empirical　risk　minimization　to　learn　causal　effects　by　adjusting　the　empirical　risks　on　the　factual　domain　and　the　counterfactual　domain　to　eliminate　the　confounding　effect　caused　by　domain　selection.　Our　method　surpasses　classic　and　cutting-edge　baselines,　as　demonstrated　through　extensive　experiments　on　three　real-world　datasets.　Furthermore,　it　achieves　unbiasedness　recommendations　effectively.　©　China　Computer　Federation　(CCF)　2024.