Human activity inference and recognition on mobile devices has many applications in personal health,wellness,and lifestyle.Offline machine learning approach with independent model learning phase and activity inference phase is a typical approach for recognition and inference,however,this offline method is not time-efficient in model updating and it cannot make full use of the latest data for model learning.In order to overcome these disadvantages,an online learning method is presented in this paper with mining multiple contextual data streams.It firstly uses current individual contextual data to infer current activity with the existed classification model and then updates the model with current data and its true activity label.Two assessments were conducted on a workstation based on a live dataset collected on the campus of Texas A&M University-Corpus Christi(TAMUCC)by a smartphone.The first assessment is a 10-fold cross-validation test to compare the performance of the offline approaches(naive Bayes offline and C4.5 algorithm)and online approaches(naive Bayes online and Hoeffding tree algorithm)in terms of classification accuracy,learning time efficiency and memory requirements.And the other assessment is an interleaved test-then-train test to demonstrate the “real-time” property of online learning method(with naive Bayes online and Hoeffding tree algorithm).The result of first assessment shows that the online learning method is more time-efficient than offline learning,and the second assessment demonstrates that models can be updated at every example which means online learning method is able to make maximum use of the latest available data.