SUMMARY
Knowledge discovery about people and cities from emerging location data has been an active research field but is still relatively unexplored. In recent years, a considerable amount of work has been developed around the use of social media data, most of which focusses on mining the content, with comparatively less attention given to the location information. Furthermore, what aggregated scale spatial patterns show still needs extensive discussion. This paper proposes a tweet-topic-function-structure framework to reveal spatial patterns from individual tweets at aggregated spatial levels, combining an unsupervised learning algorithm with spatial measures. Two-year geo-tweets collected in Greater London were analyzed as a demonstrator of the framework and as a case study. The results indicate, at a disaggregated level, that the distribution of topics possess a fair degree of spatial randomness related to tweeting behavior. When aggregating tweets by zones, the areas with the same topics form spatial clusters but of entangled urban functions. Furthermore, hierarchical clustering generates a clear spatial structure with orders of centers. Our work demonstrates that although uncertainties exist, geo-tweets should still be a useful resource for informing spatial planning, especially for the strategic planning of economic clusters.