The edge-cloud collaboration networks have been applied to support delay-sensitive Internet of Things (IoT) applications, where applications are represented in terms of service compositions. In this setting, IoT services should be configured mostly at the network edge, and they are offloaded to the cloud only when the capacity of edge nodes can hardly meet the requirement. To solve this problem, this paper proposes to configure IoT services with temporal constraints discovered from event logs. Service configuration is reduced to a constrained multi-objective optimization problem, which can be solved by an improved non-dominated sorting genetic algorithm II. Experimental results demonstrate the efficiency of this technique in comparison with baseline techniques on delay sensitivity and energy consumption.