Abstract:

In recent years，such database fields as multimedia information retrieval，similarity join and time series matching，where similarity search has attracted much attention.Existing researches mostly compute nearest neighbor to solve problems about search target set，such as kNN and kNNJ，by metric distance functions in the Euclidean space.But some studies showed that high dissimilarity dimensions had got great effect on the accuracy of answer and flexibility and robustness still were lacked in corresponding solutions.Thus centralized dynamic subspace or partial dimensions similarity search problem and algorithms were proposed at first.Furthermore，with the emerge of very large dataset，centralized algorithms can，t extend very well.Finally，the distributed ones under hadoop framework were proposed.Experiments prove that distributed algorithms outperform centralized ones in the performance without accuracy loss.

Key words: adaptive subspace, similarity search, non-metric distance method, MapReduce distributed computing framework