Algorithms to Discover Complete Frequent Episodes in Sequences

Serial episode is a type of temporal frequent pattern in sequence data. In this paper we compare the performance of serial episode discovering algorithms. Many different algorithms have been proposed to discover different types of episodes for different applications. However, it is unclear which algorithm is more efficient for discovering different types of episodes. We compare Minepi and WinMiner which discover serial episodes defined by minimal occurrence of subsequence. We find Minepi cannot discover all minimal occurrences of serial episodes as the literature, which proposed it, claimed. We also propose an algorithm Ap-epi to discover minimal occurrences of serial episode, which is a complement of Minepi. We propose an algorithm NOE-WinMiner which discovers non-overlapping episodes and compare it with an existing algorithm. Extensive experiments demonstrate that Ap-epi outperforms Minepi(fixed) when the minimum support is large and NOE-WinMiner beats the existing algorithm which discovers non-overlapping episodes with constraints between the two adjacent events.