In this paper we introduce a simple yet experimentally convincing approach in the research field of source selection for content-based similarity search in P2P networks or, more concretely, in summary-based P2P systems. In these systems, summaries are used for data source selection when performing k-NN queries on distributed collections of documents represented by feature vectors.
We introduce a new type of cluster-based summaries for source selection that can efficiently and cheaply be calculated and distributed in P2P networks. For the summaries generation, a very large number of sample points is used. Each peer in the network assigns its indexing data to their corresponding closest sample points and publishes its constructed summary. We evaluate the quality of these summaries when changing the number of sample points used in experiments on real-world image feature data obtained from a large crawl of the flickr web photo community and show that for higher numbers of sample points we achieve a better retrieval performance. Our experiments show that the proposed summaries yield four times better performance with respect to previous methods. Intuitively, there are some disadvantages to this approach due to the large size of the generated summaries. We show experimentally, that these disadvantages can easily be overcome due to the sparse nature of the generated summaries by simple compression techniques.