Recently,　recommender　system　suffers　extremely　from　both　interaction　bias　and　sparsity.　The　conventional　unified　embedding　learning　policies　fail　to　consider　the　imbalanced　interaction　issue　and　produce　suboptimal　representations　of　users　and　items　for　recommendation.　Towards　the　end,　this　work　dedicates　to　bias-aware　embedding　learning　in　a　decomposed　manner　and　proposes　a　Counterfactual　Embedding　Learning　(CEL)　for　debiased　recommendation.　Instead　of　debiasing　with　sampling　uniform　interactions,　we　follow　and　capitalize　the　natural　bias　distribution　to　model　users　and　items　with　a　counterfactual　hypothesis.　Concretely,　embeddings　of　users　and　items　are　built　with　common　and　special　causes　to　model　latent　frequent　and　infrequent　bias　isometrically.　Particularly,　relying　on　the　frequency　gap　between　common　and　special　causes,　CEL　introduces　a　bias-aware　counterfactual　masking　on　interactions　and　forms　multiple　counterfactual　worlds　to　extract　effects　of　various　special　causes　comparing　with　the　factual　world.　Experiments　show　that　the　proposed　CEL　performs　superior　to　state-of-the-art　baselines.