The problem of discovering regions that support particular functionalities in an urban setting has been approached in literature using two general methodologies: top-down, encoding expert knowledge on urban planning and design and discovering regions that conform to that knowledge; and bottom-up, using data to train machine learning models, which can discover similar regions. Both methodologies face limitations, with knowledge-based approaches being criticized for scalability and transferability issues and data-driven approaches for lacking interpretability and depending heavily on data quality. To mitigate these disadvantages, we propose a novel framework that fuses a knowledge-based approach using design patterns and a data-driven approach using latent Dirichlet allocation (LDA) topic modeling in three different ways: Functional regions discovered using either approach are evaluated against each other to identify cases of significant agreement or disagreement; knowledge from patterns is used to adjust topic probabilities in the learning model; and topic probabilities are used to adjust pattern-based results. The proposed methodologies are demonstrated through the use case of identifying shopping-related regions in the Los Angeles metropolitan area. Results show that the combination of pattern-based discovery and topic modeling extraction helps uncover discrepancies between the two approaches and smooth inaccuracies caused by the limitations of each approach.