Autonomous computational creativity systems must not only have the ability to generate artifacts, but also to select the best ones on the basis of some assessment of quality (and novelty). Such quality functions are typically directly encoded using domain knowledge or learned through supervised learning algorithms using labeled training data. Here we introduce the notion of unsupervised computational creativity; we specifically consider the possibility of unsupervised assessment for a given context by generalizing artifact relationships learned across all contexts. A particular approach that uses a knowledge graph for generalizing rules from an inspiration set of artifacts is demonstrated through a detailed example of computational creativity for causal associations in civic life, drawing on an event dataset from political science. Such a system may be used by analysts to help imagine future worlds.