【摘要】 当前时间序列Shapelet发现算法普遍采用穷举法，需要计算所有时间序列子序列的信息增益，效率较低。针对此问题，提出一种基于矩阵轮廓的Shapelet发现算法。选出最具代表性的时间序列对，计算其轮廓矩阵和差异向量，找到一簇关键区域；对找到的关键区域进行剪枝；在关键区域上搜索Shapelet并计算其信息增益，提升算法效率。在15个UCR数据集上，通过时间序列二分类实验对所提Shapelet发现算法进行验证。实验结果表明，所提算法结合Shapelet转换后具有较强分类能力，计算效率明显优于现有Shapelet发现算法。更多还原

【Abstract】 Current algorithms of time series Shapelet discovery, which are widely based on enumeration, require to figure out the information gain of time series subsequence and suffer from low efficiency. To handle this problem, an algorithm of Shapelet discovery based on matrix profile was proposed. The pairs of time series that were representative were selected out, their matrix profile and difference vector were calculated, and a set of critical areas was found. The critical areas were pruned. The Shapelet in the critical areas was searched and the information gain was calculated, improving the efficiency of the algorithm. Experiments of binary classification of time series were performed to evaluate the proposed Shapelet discovery algorithm. The results show that the proposed algorithm combined with Shapelet transformation has strong classification ability, and its efficiency is obviously superior to the existing algorithm of Shapelet discovery.更多还原