Causal feature selection for physical sensing data

Abstract

Identifying the causal features from multi-dimensional physical data streams is one of the underpinnings for the success of the data-driven inference tasks. Prior research utilizes the 'correlation' or 'mutual information' to select features, which ignored the crucial inherent causality behind the data. In this work, we consider the problem of selecting causal features from streaming physical sensing data. Inspired by a metric from information theory which calibrates both instantaneous and temporal relations, we formulate the causal feature selection as a cardinality constrained joint directed information maximization (i.e., Max-JDI) problem. Then we propose a near optimal greedy algorithm for streaming feature selection and present an information-interpretive solution for the cardinality constraint presetting. The proposed method is evaluated on a real-world case study involving feature selection for the power distribution network event detection. Compared with other selection baselines, the proposed method increases the detection accuracy by around 5%, while concurrently reduces the computation time from several weeks to within a minute. The promising results demonstrate that it can be applied to optimize the energy operation and enhance the resilience of power buildings.