The recognition of individual activities helps in the realisation of functions such as user profiling, personalized recommendations, abnormal behaviour detection, city-wide group behaviour analysis and resource allocation optimisation. A recognition method for the semantics of individual activities based on sparse social media check-in data was proposed. The temporal periodicity and tendency features of activity behaviors were extracted from the check-in data, and a spatial preference quantification algorithm was utilized to extract the preferences of groups and individuals from the spatial relevance between individual and group activities. The natural language embedding model BERT was used to extract the semantics of POIs (point of interest). The temporal features, spatial preference features and text features of POI′s names constituted the joint spatio-temporal features characterizing group and individual preferences, and the joint features were classified by the extreme gradient boosting classifier to obtain the activity semantic recognition results. With the results of comparison experiments and ablation experiments on the Foursquare dataset, it was validated that the model proposed can effectively improve the accuracy of activity semantics recognition.