Mining frequent pyramid patterns from time series transaction data with custom constraints

Abstract

For the problem of mining pyramid scheme patterns, the traditional sequential pattern mining algorithm Prefixspan has many disadvantages such as poor timeliness, uniform threshold, etc. Therefore, we propose a timeliness variable threshold and increment Prefixspan algorithm, named TVI-Prefixspan, for mining the sequential patterns from time series transaction data. To be specific, TVI-Prefixspan aims to mine the patterns that co-occurrence in both an individual sequence and different sequences with high frequency. The most important challenges are how to define the thresholds of frequent one-item and pyramid patterns. We firstly analyze the attributes of the patterns which are hidden in the financial activities between different bank accounts. Secondly, the frequent threshold of each one-item is determined by its different frequency value in normal and pyramid related transaction sequences. We also consider the special relationships in both numerical values and time-series aspects between each pattern’s item. Therefore, TVI-Prefixspan produces the frequent one-item set based on its difference of the normal frequency, and then, mines the pyramid patterns with formulated relation constraints. For describing the correlation, we consider sequential, time interval and one-off constraints simultaneously. The experimental results, in real financial data containing pyramid transactions, show that TVI-Prefixspan algorithm succeeds in mining pyramid scheme patterns quickly and effectively. It is superior to traditional sequential pattern mining algorithms such as Prefixspan in efficiency and mining effect.