SWAM: a family of access methods for similarity-search in peer-to-peer data networks

Reviewer: Maytham Hassan Safar

Banaei-Kashani and Shahabi describe querical data networks (QDNs), which are large-scale, self-organizing, distributed query processing systems. As an example, they present a peer-to-peer network. The data is distributed among the nodes, and each node knows its data, plus the data of some of the nodes that it is connected to. To keep the network scalable, there are only a few connections made from a node to other nodes. The authors provide a method to perform typical similarity queries on peer-to-peer networks. In addition, they provide a mechanism to support different query types, such as range, and k-nn queries. The authors' solution relies on already existing data partitioning techniques in the value space, where each data item is considered to exist only once. The space of all the data items in the network is decomposed once hierarchically, and once using Voronoi diagram partitioning techniques. Then, a grid of connections is created that consists of some random connections between nodes, and connections with other nodes that are considered as neighbors of a node by the decomposition method. The two components are overlaid, and the resulting grid forms a connection network of the "virtual" nodes of data items. Finally, a reverse mapping, from each data item in this virtual space to the actual storage nodes, is performed, and the resulting network is presented. Queries are shown to be exactly answerable using the resulting network. The authors fail to describe the applications that would use their technique with examples of real queries. Instead, only abstract configurations are mentioned in their experimental results. The curse of dimensionality issue, which implies that, in high-dimensional spaces, there is no useful locality to exploit, is not discussed. This locality is crucial to the proposed method. The authors should more carefully investigate what happens to the communication cost as the dimensionality of their space increases. The authors assume exactly one item per node, and provided an argument showing how the general case can be transformed to this restricted case. By doing this transformation, and making the data size proportional to the number of nodes in the system, one loses track of data dependence. The authors imply that the complexity of their method is linear in the data set size, regardless of the number of dimensions of the data, which is a strange-sounding result. Overall, this is a good work, except for the experimental setup. Online Computing Reviews Service