Discovering Mass Activities Using Anomalies In Individual Mobility Motifs

The development of mobile communication technology and location technology allows researchers to obtain a large number of human mobility data and capture trajectory of users.These trajectory data with user behavior characteristics can be drawn a lot of meaningful conclusions through data mining.In past works,researchers have often focused on the more common human mobility patterns,which can reveal the general laws of group behavior.At the same time,some of the less common mobility patterns with long tail distributions are often overlooked by researchers.Mass activities can cause a group to perform abnormal mobility patterns.By mining the occurrence regularity and distribution of these abnormal patterns,is it possible to detect large-scale activities in reverse?This paper proposes a mass activity recognition model based on individual mobility pattern anomaly.It can use the impact of mass activities towards group movement to complete large-scale activity detection on the user trajectory data set.In our research,firstly,the user movement trajectory is generated by data fusion based on WIFI data and e-card transaction data obtained from a university campus.After semi-supervised clustering for user selection,semantic clustering and module extraction algorithms are used to construct a mobility motif set for each user.According to the statistical analysis results of the set,we can get abnormal mobility motif distribution,and the abnormal motif feature is extracted from the distribution.Then with additional features added,the training data set is accomplished.We use cross-validation to organize data set and undersampling and threshold shift method are used to solve the problem of sample imbalance.Then,an improved Gradient Boosting Decision Tree(GBDT)algorithm is used to predict whether each building will perform mass activities every day.Finally,the prediction results of our model and the conventional clustering algorithm are compared,which proves that our model has better prediction effect.On the one hand,the study of the mobility patterns of human groups has been carried out,which further reveals the influence of human mobility and external factors.On the other hand,this kind of large-scale clustering activity detection is also important for the management of new public safety events.