Research On The Key Technologies For Temporal Event Sequence Data Visualization

Temporal event sequence data is a dataset organized by a succession of events in chronological order.It records the activities of human beings or other entities in the real world,such as mobile phone operation records,web page click records and communication records.The analysis of temporal event sequence data aims to discover the activity patterns of various entities,such as periodic patterns,frequent patterns,similar patterns and aberrant phenomena.Many machine learning and data mining methods are applied to mine entity activity patterns automatically.However,in some application scenarios,automated methods may sever the organic connection between people,systems and data.Users may have difficulty understanding or trusting the pattern mining results learned from these automated methods.Visualization technology can present the temporal event sequence data in the form of graphic images.Leveraging human visual perceptions,combining automatic methods and interactions,it can help users explore entity activity pattern interactively.This thesis takes the time dimension,event dimension and entity dimension in the temporal event sequence data as the research perspective.It focuses on the visual presentation and interaction analysis methods of temporal event sequence data with multi-entities and single-event,singleentity and multi-events,as well as multi-entities and multi-events.The specific contents are as follows:(1)An edge-based and multi-class blue noise sampling based algorithm,which effectively improves the overall readability of the massive sequence view.The massive sequence view is a technique commonly used to visualize the dynamic changes of multi-entities’behaviors and relationships.With the increasing of data volume,the overall readability of this view deteriorates due to edge overlap,which results in visual confusion easily.This thesis proposes an edge-based and multi-class blue noise sampling based algorithm.This algorithm redefines the sampling distances,and the conflict check process in the traditional multiclass blue noise sampling method,and adopts a partition sampling strategy.It can solve the three defects of instability,imbalance and loss of outliers in the Monte Carlo-based edge sampling method,alleviate the visual confusion and improve the readability of the massive sequence view.(2)An interaction analysis method for frequent pattern exploration of temporal event sequence data,which can effectively analyze the characteristics of events in the sequential and relational space.Frequent patterns in long time sequences with multi-events may reflect the correlation pattern and causality relationship between events,and produce summarization for dynamic patterns of events in the time dimension.This thesis proposes an interaction analysis method for frequent pattern exploration,in the context of the mobile phone operation behavior sequences of smart phone users.This method measures the temporal randomness and correlation of frequent events,by using the information entropy and Apriori algorithm respectively.We also propose a visual analysis system,which interactively visualizes and illustrates the correlation between events.(3)An interaction analysis method for identifying and analyzing similar temporal event sequences,which can effectively identify and analyze similar sequences in multi-levels and multi-granularities.This method extracts composite events to generalize multiple original events,and simplifies the original event sequences based on the composite events.By doing so,it can reduce the diversity of event sequences and summarize the evolution patterns of event sequences.We also propose a visual analysis system,which interactively visualizes and illustrates the critical evolution patterns that help define the common characteristics of each sequence groups and reveal differences between different sequence groups.