Nonoccurring Sequential Behavior Analytics

Publication Type:
Issue Date:
Full metadata record
Behavior analytics has attracted increasing attention in broad communities as a major research area in understanding and managing the dynamics of complex systems and problems such as series of medical treatments, interactions between customers and service providers, and online communications. Sequential behavior analytics aims to understand, analyze, detect, and predict existing or future behaviors and behavior sequences. Existing methods for sequential behavior analytics only focus on occurred or to-occur behaviors (also called positive behaviors), while ignoring nonoccurring behaviors (also called negative behaviors), which are often useful for understanding, managing and predicting hidden or unseen yet important behaviors that differ from and typically mix with occurred ones. Nonoccurring behaviors complement occurring ones for complete and deep behavior analytics, while very limited theoretical progress has been made. This thesis studied the theory to comprehensively model the complex relations within and between behaviors and to effectively discover and predict interesting sequential occurring and nonocurring behaviors. Specifically, it focused on (1) forming a comprehensive and systematic representation, formalization, and theoretical system for defining and representing the concepts, problems, constraint settings, and negative containment of nonoccurring sequential behavior (NSB) analytics; (2) efficiently discovering the high-frequency negative sequential patterns (NSP) composed of both occurring/nonoccurring behaviors; (3) discovering the representative NSP subset by exploring the complicated explicit/implicit behavior relations; and (4) enabling the sequential basket recommendation system (SBRS) through learning behavior relations and interactive feedback. Accordingly, this thesis proposed (1) a vertical NSP mining framework and its instantiation for the efficient discovery of the complete set of NSP with the loose negative element constraint via the vertical representation of each sequence, which guarantees the coverage of flexible patterns with complicated behavior relations; (2) a determinantal point processes-based (DPP-based) representative NSP discovery approach for the selection of a representative subset of the high-quality and diverse patterns by jointly modeling explicit and implicit sequential element/pattern relations; and (3) a hierarchical attentive encoder-decoder model for interactive sequential basket recommendation, which jointly models both intra-/inter-basket relations in sequential user basket behaviors as well as incorporates positive/negative feedback to enable negative feedback-based refinement. The extensive empirical analysis of the proposed methods demonstrated that our methods performed significantly better than the state-of-the-art methods in the same domain in terms of multiple evaluation metrics.
Please use this identifier to cite or link to this item: