Behavior refers to the action or property of an actor, entity or otherwise, to situations or stimuli in its environment. The in-depth analysis of behaviour has been increasingly recognized as a crucial means for understanding and disclosing interior driving forces and intrinsic cause-effects on business and social applications, including web community analysis, counterterrorism, fraud detection and customer relationship management, etc. Currently, behavior modeling and analysis have been extensively investigated by researchers in different disciplines, e.g. psychology, economics, mathematics, engineering and information science. From those diverse perspectives, there are widespread and long-standing explorations on behavior studies, such as behavior recognition, reasoning about action, interactive process modeling, multivariate time series analysis, and outlier mining of trading behaviors.
All the above emerging methods however suffer from the following common issues and problems to different extents: (1) Existing behavior modelling approaches have too many styles and forms according to distinct situations, which is troublesome for cross-discipline researchers to follow. (2) Traditional behavior analysis relies on implicit behavior and explicit business appearance, often leading to ineffective and limited understanding on business and social activities. (3) Complex coupling relationships between behaviors are often ignored or only weakly addressed, which fails to provide a complete understanding of the underlying problems and their comprehensive solutions. (4) Current research usually overlooks the checking of behavior interactions, which weakens the soundness and robustness of models built for complex behavior applications. (5) Most of the classic mining and learning algorithms follow the fundamental assumption of independent and identical distribution (i.e. IIDness), but this is too strong to match the reality and complexities in practical applications.
With the deepening and widening of social/business intelligences and their networking, the concept of behavior is in great demand to be consolidated and formalized to deeply scrutinize the native behavior intention, lifecycle and impact on complex problems and business issues. In the real-world applications, group behavior interactions (i.e. coupled behaviors) are widely seen in natural, social and artificial behavior-related problems. The verification of behavior modeling is further desired to assure the reliability and stability. In addition, complex behavior and social applications often exhibit strong explicit or implicit coupling relationships both between their entities and properties. They can not be abstracted or weakened to the extent of satisfying the IIDness assumption. These characteristics greatly challenge the current behavior-related analysis approaches. Moreover, it is also very difficult to model, analyze and check behaviors coupled with one another due to the complexity from data, domain, context and impact perspectives.
Based on the above research limitations and challenges, this thesis reports state-of-the-art advances and our research innovations in modeling, analysing and learning coupled behaviors, which constitute the coupled behaviour informatics. Coupled behaviors are categorized as qualitative coupled behaviors and quantitative coupled behaviors, depending on whether the behaviour involved is qualified by actions or quantified by properties.
In terms of the qualitative coupled behavior modeling and analysis, we propose an Ontology-based Qualitative Coupled Behavior Modeling and Checking (OntoB for short) system to explicitly represent and verify complex behaviour relationships, aggregations and constraints. The effectiveness of OntoB system in modeling multi-robot behaviors and their interactions in the Robocup soccer competition game has been demonstrated.
With regard to the quantitative coupled behavior analysis and learning, we carry out explorations on three tasks below. They are under the non-IIDness assumption of entities or properties or both of them, which caters for the intrinsic essence of real-world problems and applications.
For numerical coupled behavior analysis, we introduce a framework to address the comprehensive dependency among continuous properties. Substantial experiments show that the coupled representation can effectively model the global couplings of numerical properties and outperforms the traditional way. For categorical coupled behavior analysis, we present an efficient data-driven similarity learning approach that generates a coupled property similarity measure for nominal entities. Intensive empirical studies witness that the coupled property similarity can appropriately quantify the intrinsic and global interactions within and between categorical properties for especially large-scale behavior data. For coupled behavior ensemble learning, we explicate the couplings between methods and between entities in the application of clustering ensembles, and put forward a framework for coupled clustering ensembles (CCE). The CCE is experimentally exhibited to capture the implicit relationships of base clusterings and entities with higher clustering accuracy, stability and robustness, compared to existing techniques. All these models and frameworks are supported by statistical analysis.
Finally, we provide a consolidated understanding of coupled behaviors by summarizing the qualitative and quantitative aspects, extract the multi-level couplings embedded in them, and then formalize a coupled behavior algebra at its preliminary stage. Many open research issues and opportunities related to our proposed approaches and this novel algebra are discussed accordingly.
Under varying backgrounds and scenarios, our proposed systems, algorithms and frameworks for the coupled behavior informatics are evidenced to outperform state-of-the-art methods via theoretical analysis or empirical studies or both of them. All these outcomes have been accepted by top conferences, and the follow-up work has also been recognized. Therefore, coupled behavior informatics is a promising though wholly new research topic with lots of attractive opportunities for further exploration and development.