Interactive Mining And Visual Analysis Of Large-scale User Behavior Data

Millions of users from different geolocations are consuming, producing, and disseminating huge volumes of highly diverse information, ranging from social media data in golobal level to urban level residents travel data. With the rapid development and prevalence of mobile technologies, the data generated by users will continue to increase and have more diversity. Beisdes, the data will be more easily to be acquired than before. The analysi and understanding of user behavior data with respect to the patterns in temporoal and spatial dimension poses presents many useful applications, especially in fields of commerce, media, government and academic study.However, understanding user behavior data is hindered by following major challenges. First, user behavior data is of multi-dimension and multi level, which usually involves time, space and other dimension such as text content. Analysis with respect to different spatial and temporal granularity/scale often leads to different results. Second, the patterns in user behavior data usually evolves with time and space dynamically, and different users may interact with each other. Third, the user behavioral data involves huge amount of users and records, thus, how to stably and effectively store and analyze the data is still questionable. Last but not least, the complex, multi-dimensional, and multi level dynamic user behavior data is nontrivial to present visually and track timely.Visual analytics employs interactive visualizations to integrate users’ knowledge and inference capability into numerical/algorithmic data analysis processes, and assist users in mining and presenting patterns in huge volumes of data. This dissertation aims to employ the method of visual analysis to study the user behavior pattern in temporal, spatial and spatio-temporal dimension. We study and discuss different issues including user behavior data modelling, visualization design, development of interactive visual analysis system, and visualization evaluation. The main contributions of this dissertation are as follows:We propose a new model that can quantitatively characterize the dynamic topic coopetitionrelated interactions as well as the influence of topic leaders on such interactions. We model the complex interactions among topics as a combination of carry-over, coopetition recruitment, and coopetition distraction effects. We also design Evo River, a time-based visualization that allows users to explore coopetition-related interactions and to detect dynamically evolving patterns as well as their major causes.We propose a dynamic social gravity model to quantify the time-varying spatial interaction behavior among social media users in information diffusion. The dynamic social gravity model includes three factors that are theoretically significant to the spatial diffusion of information: geographic distance, cultural proximity, and linguistic similarity. Temporal dimension is also taken into account to help us detect recency effect. Furthermore, Social Flow, a visual analytic system, is developed to support both spatial and temporal investigative tasks.We present a novel visualization that can seamlessly embed temporal displays into a map for occlusion-free visualization of both the spatial and temporal attributes of the data. We first extend the seam carving algorithm to broaden the roads of interest in a map with the least distortion to other areas, and then embed temporal displays into the roads to reveal temporal patterns without the occlusion of map information. We study various design choices in our method, including the encoding of the time direction and temporal display, and conduct two comprehensive user studies to validate our design decisions.We evaluate above methods with real world data including social media data and traffic flow data in a major city, and offer profound insights into the dynamics of topic coopetition, the spatiotemporal diffusion process among different users, and the urban residents travel behavior. We further evaluate the effectiveness and usefulness of above methods with user study and domain experts.