A transfer learning paradigm for spatial networks

Advances in machine learning and the availability of spatial data have seen remarkable improvements in recent times. This parallel growth has influenced the increased application of traditional data mining techniques for knowledge discovery on spatial data. However, these techniques assume that the data is drawn from an independent and identical distribution whereas spatial data is inherently dependent and heterogeneous. This contradiction strongly suggests that a crass application of conventional data mining techniques to spatial data would be suboptimal. In this paper, we evaluate the relatedness of street networks using a transfer learning methodology within the formal contexts of spatial data. Adopting a statistical multi-measure, we analyze street networks from eight cities in an attempt to ascertain their similarities. We predict the street types using random forests and evaluate the accuracies as a function of transfer polarity. Positive transfer is when the transferred models perform better than the parent model or negative transfer when it is worse. With an overall average accuracy of 85%, our results show that it is possible to generalize machine learning models onto different domains and still produce excellent results. Also, we demonstrate that the improved or loss of model accuracy can be explained by the proportion of statistical similarity between the domains. This observation confirms that a measure of inter-domain similarity solely based on geo-political boundaries will be erroneous. The techniques we have described are a statistically sound foundation for analysis of similarities in the spatial context. It can be adopted towards understanding the extent of model generalization for spatial networks.