We are developing an interactive service robot which provides personal greetings to customers, using a machine-learning approach based on observations of a customer's appearance or behavior from on-board or environmental sensors. For each visit, several features are recorded, such as 'time of day' or 'number of people in group.' A set of classifiers trained by human coders compare the current features with the person's individual history, to determine an appropriate feature for a robot to speak about. This system enables the robot to make context-appropriate comments such as 'good morning, you're here very early today.' We present the design of our system and an encouraging set of preliminary prediction results based on one month of data taken from real customers at a shopping mall. © 2013 IEEE.