As an important category of social communication,facial expressions provide abundant social and emotional information.However,it has been a challenge for computer systems to provide real time accurate recognition of facial expressions from multimedia content and make the analysis interpretable.To address these issues,we propose a sparse tagging-like approach to jointly learn Action Units for facial expressions recognition,which can be utilized for social interaction analysis.Specifically,we regard the recognition of the combination of Action Units as tagging images.Under this approach,the computational complexity is substantially reduced because only matrix multiplications are involved.In order to make the analysis interpretable,we introduce a sparse term into our approach to reinforce the sparseness of the combination of Action Units.Experiments in four benchmark datasets demonstrate that,compared with existing algorithms,our method achieves much faster speed,and higher interpretability and robustness,while yielding a matching accuracy.