Qualitative-probabilistic-network-based modeling of temporal causalities and its application to feedback loop identification

Modeling and analysis of time-series data attract much attention in data mining and knowledge discovery community due to their many applications in financial analysis, automation control, etc. In such applications, time-series data usually contain several attributes that may be causally dependent in historical time slices. Dangerous feedback loops of attributes’ dependent relationships can make the system collapse due to amplification or oscillation of attribute values. Motivated by efficient analysis of causalities in time-series data, we propose a temporal qualitative probabilistic graphical model in this paper. From given time-series sample data, we construct the structure of the temporal qualitative probabilistic network (TQPN) and derive the corresponding qualitative influences on directed edges. We then present the approach for TQPN reasoning with time-series features. Consequently, positive time-series feedback loops are defined, and the approach to identify them is proposed. Preliminary experiments show that our proposed method is not only feasible but also efficient.