Semantic-aware Temporal Data Mining

Temporal data mining is a hot research topic.It has wide applications,such as realtime POI(Point of Interest)recommendation,risk prediction in financial systems,scholar archiving,and retrieval.Compared with static data,temporal data not only carries timetagged information but also contains the dynamic evolution of data over time.There are two main problems in recent research: firstly,most of the research work only aims at the characterization of data at each moment and predict the future state but neglects the modeling of the dynamic evolution process;secondly,some time-series data are very sparse and contain noise,current algorithms often have difficulty in capturing high-level semantic features and are not sufficiently explainable.To address these problems,the main contributions of this thesis are as follows:In order to solve the problem of sparsity in POI recommendation data,this thesis proposes an algorithm DSPR that can model and characterize users’ decisions to access POI from their real-time demands.DSPR uses additional semantic information to reduce the impact of POI data sparsity on traditional recommendation algorithms and models and characterizes users’ real-time needs to access POI.External experiments show that the DSPR outperforms state-of-the-art POI recommendation algorithms and can learn the user’s real-time demands with good interpretation.In the second study,to fill the current gap in modeling and to characterize the data evolution process,this thesis proposes an event representation-based algorithm,LDBR,for node behavior modeling and learning.The LDBR algorithm abstracts the node behavior’s evolution process in temporal collaboration network into node behavior and partitions the node behavior into a series of consecutive events.LDBR learns the evolution of nodes by learning the event representation of nodes.The proposed LDBR algorithm can distinguish the semantics of nodes in real-time and accomplish the temporal link prediction task.External experiments show that the LDBR algorithm achieves better link prediction accuracy than the traditional dynamic network characterization algorithm.