Research On Key Technologies Of Trajectory Data Mining

The advances in location-acquisition techniques and location-based social networks generated trajectory data at an explosive rate.Trajectory data contains spatiotemporal fea-tures and mobility behavior pattern of the moving object.Knowledge discovery and data mining on trajectory data is a research subject that it has great theory values as well as practical significance.Trajectory data mining is the orientation of future urban computing and smart city,which has wide applications such as traffic planning and control,urban emergency service,location prediction and recommendation,crowd mobility sensing,so-cial behavior analysis.However,trajectory data mining is still facing some problems in realistic scenarios.First,in the absence of semantics of trajectory data,it is difficult to comprehend the implicit semantic features in trajectory data.Second,the representation method of a sequence of trajectory points is not suitable for massive data mining and ma-chine learning models.Then,the performances of existing location prediction methods are insufficient for practical needs.Therefore,this dissertation focus on three key tech-nologies such as trajectory data semantics inference,trajectory representational learning,improving the performance of location prediction.The main contributions of this disser-tation are as follow:(1)For the problem of lack of trajectory data semantics,a novel location semantics inference with graph convolutional networks(SI-GCN)is proposed in this dissertation.Different from established approaches manually extracted location features or spatiotem-poral patterns,SI-GCN is allowed to automatically extract location spatial and temporal features via network representation learning and variational autoencoder,respectively,and then SI-GCN builds spatiotemporal features.SI-GCN leverages graph convolutional net-works to capture high-order visited relations in a user-location bipartite network.Mean-while,SI-GCN introduces a self-attention mechanism to learn different contributions of neighbor nodes in the user-location bipartite network.Extensive experimental results on two check-in data sets indicate SI-GCN outperforms the state-of-art methods.(2)With a view to tackling the problem of difficulty in trajectory representation,a novel space-time architecture for semantic trajectory representation(STAR)is proposed in this dissertation.Existing representation methods tend to ignore the temporal information carried in trajectories,or treat spatial and temporal information separately.Meanwhile,existing representation methods are not suitable for wide data mining and machine learning models.STAR builds a unified framework for learning spatial and temporal information in trajectory,which adopts distributed vector representation and adaptive Hawkes process.To enrich trajectory context,STAR extends geometric context,trajectory context and se-mantic context as context information.STAR aims to learn trajectory representation as a bridge to connect location and temporal information as a whole,by predicting next visiting location and time based on distributed vector representation and adaptive Hawkes process.Extensive experiments on several real-world trajectory data sets show that STAR outper-forms existing representation methods in the field of semantic similarity query,similarity measurement and anomaly detection.(3)Aimed at the problem of bad performances of location prediction,which caused by high randomness and sparse in the individual moving pattern.A new robust location prediction model that considers both individual preferences and social interactions(PSI)is proposed in this dissertation.PSI introduces group level mobility patterns to alleviate the effect of randomness and sparsity for improving location prediction performances.To characterize exterior social interactions,an associated group is identified,and an outline of the group moving patterns is then extracted based on association rule mining.Finally,the next location is predicted by learning the individual's regular patterns and group moving patterns via a pair-wise ridge regression.Building upon group-level pattern mining,PSI provides a more robust prediction model by learning both individual and group trend in-formation simultaneously,alleviating the randomness and sparsity of location prediction from individual historical trajectory data only.In contrast to the traditional approaches,PSI achieves a better prediction performance compared to the state-of-the-art methods.(4)In allusion to the problem of insufficient study on the factor and dynamic mo-bility mechanism of a user's check-in behavior,a dynamic model jointly performs geo-aware personal preference learning and social interaction excitation modeling(DGPS)is proposed in this dissertation,which models evolving user's check-in behavior.DGPS constructs geo-aware location features(including semantic features,latent features and dynamic features)to represent the evolution of geographical location influence and pro-vide deep insights into user preferences.DGPS leverages a temporal point process that is used to capture the context dependency and dynamic changes in users' check-in activities.Furthermore,geo-aware preferences learning of a user and social interaction excitations are embedded into the conditional intensity function of temporal point process,which aims to learn driving factors of user's check-in behavior.Comprehensive experiments results on several real-world check-in data sets indicate DGPS outperforms existing state-of-art approaches.