Mining Frequent Patterns with Temporal Effect: A Case of Accident Path Analysis

Abstract

Frequent item-set generation is usually applied on large databases to mine item-sets recurring together. The objective of such analysis is to identify patterns in dataset which are of managerial interests. In cases where the dataset is spread over significant period of time, the results can sometime be misleading. The reasons are twofold: (i) pattern existing with high frequency at certain time period and later experiencing decreasing trend will appear in result and (ii) pattern present in recent time periods with increasing trend but not with significantly high frequency will not appear in result. Due to this the results become bias to time effect. In the proposed work, we have included the time effect in data during pattern mining by exponentially weighing the dataset i.e., the more recent the data, higher the weight assigned to it. The method is applied to accident path components from incident records of an integrated steel plant. The results from the proposed work and traditional frequent item-set mining are compared and results are inferred as: (i) the patterns present only in traditional method’s result are the accident patterns which are under control and have effective risk control system (RCS), (ii) patterns present in both the results are paths where either RCS is absent or ineffective, and (iii) the patterns present only in proposed method’s result will have higher probability to recur in future. The proposed method provides new direction in interpretation of pattern mining from large dataset of long period.