Summary: | 博士 === 臺灣大學 === 資訊管理學研究所 === 98 === In this dissertation, we propose a novel data representation and three algorithms, MaxFSP, PCP, and CVP algorithms. The MaxFSP algorithm is proposed to find the maximal frequent spatial patterns in image databases. The PCP algorithm is designed to find the frequent closed spatial patterns in image databases. The CVP algorithm is developed to find the frequent closed spatial-temporal patterns in video databases. In the MaxFSP and PCP algorithms, in addition to using the anti-monotone pruning strategy to prune unnecessary candidate patterns, we utilize the characteristics of the spatial relationships to design the pruning strategies. Specifically, we design pruning strategies based on the MAFIA and CHARM algorithms respectively to prune non-maximal and non-closed candidates. By using these strategies, we can prune a large number of unnecessary candidate patterns. The CVP algorithm is further considering temporal information to mine all frequent closed spatial-temporal patterns in video databases. In the CVP algorithm, we not only use the spatial relationship to prune unnecessary candidate patterns, we also propose a fast-grow pruning strategy, which can speed up the mining process in the temporal dimension. Therefore, the CVP algorithm can effectively prune the unnecessary branches in the frequent pattern tree and avoid the costly candidates’ generation. The experimental results show that the MaxFSP algorithm outperforms the modified MAFIA algorithm; the PCP algorithm outperform the modified Apriori, SASMiner, and MaxGeo algorithms; and the CVP algorithm outperform the modified Apriori algorithm
|