Recommender systems support online customers by suggesting products and services of likely interest to them. Taking into consideration the context of customers is believed to produce better recommendations, yet it poses unique challenges. If a recommendation is generated through previous ratings, narrowing down the set of ratings to those under the target context will limit the number, producing poor-quality recommendations. A common approach to improving the quality of recommendations is to aggregate ratings from a number of similar context segments; however, establishing which segments to aggregate by unguided enumerations is too computationally intensive. In this paper, we propose a novel context similarity metric to guide the aggregation process and show how it can be extended across multiple context dimensions. The metric underpins another contribution: a guided aggregation approach to context-based recommendation. This approach can be combined with traditional recommendation algorithms to improve their prediction accuracy through guided selection and inclusion of data segments for training of prediction models. We demonstrate the effect of our approach on the prediction accuracy of a popular memory-based collaborative filtering algorithm. The metric and the approach are validated using a set of data on hotel service ratings under different contexts. Four sets of validating experiments demonstrate the effectiveness of the approach. © 2013 M.E. Sharpe, Inc.