GANDIVA: Temporal Pattern Tree for Similarity Profiled Association Mining

1. Introduction

Data mining deals with the discovery of useful data patterns from enormous, massive amounts of data (Nong, 2013). Data patterns discovered by applying data mining algorithms can be classified into six major types. These include a) prediction and classification patterns b) data reduction patterns c) cluster patterns d) association patterns e) anomaly patterns f) sequential and temporal patterns. Research on temporal patterns is gaining massive importance and is quickly emerging with various important applications associated with the discovery of temporal patterns. Time profiled temporal pattern mining has various applications ranging from financial analysis to healthcare. Determining and unearthing similar temporal patterns in time-stamped temporal databases is a complex and challenging task (Y. C. Chen, 2015) (Y. C. Chen, 2016). The pioneering research (Soung, 2008) (Soung, 2009) (Soung, 2012) addressed retrieval and discovery of similar temporal association patterns by introducing support estimation and pruning approaches to reduce support computations and minimize computational complexity. However, the approach to overcome the necessity of maintaining the temporal database in the memory is not addressed till today (Shadi, 2018) (Vangipuram, 2018). Also, research by (Soung, 2008) (Soung, 2009) (Soung, 2012) did not address new distance measures for time profiled association mining. Some of the previous studies (Vangipuram, 2015) (Vangipuram, 2016) (Vangipuram, 2017) (Vangipuram, 2018) addressed new distance measures SRIHASS (used in Z-SPAMINE), ASTRA (used in MASTER), and G-SPAMINE (Shadi, 2017) that uses fuzzy distance measure. The major limitation of all the above-mentioned approaches is the inevitable requirement for the database to be present in the main memory. The research embodied in this paper aims to address “…the possibility of eliminating the need for original database retention in the main memory…” Our approach has the major advantage that the original database need not be present in the memory for future computation process. To the best of our knowledge, there is no research till date w.r.t similarity profiled temporal association mining that has proposed a tree-based approach that performed the lattice-based scan. Thus, our approach is the pioneering work in this direction. We propose a new tree structure called “GANDIVA - TIME PROFILED TEMPORAL PATTERN TREE” which eliminates the huge computational overhead involved in finding supports of itemsets.