With the availability of diverse data reflecting people’s behavior, behavior analysis has been studied extensively. Detecting anom-alies can improve the monitoring and understanding of the objects’ (e.g., people’s) behavior. This work considers the situation where objects behave significantly differently from their previous (past) similar objects. We call this locally anomalous behavior change. Locally anomalous behavior change detection is relevant to various practical applications, e.g., detecting elderly people with abnormal behavior. In this paper, making use of objects, behavior and their associated attributes as well as the relations between them, we propose a behavior information sequence (BIS) constructed from behavior data, and design a novel graph information propagation autoencoder framework called LOCATE (locally anomalous behavior change detection), to detect the anomalies involving the locally anomalous behavior change in the BIS. Two real-world datasets were used to assess the performance of LOCATE. Experimental results demonstrated that LOCATE is effective in detecting locally anomalous behavior change.