Mining recent frequent cooccurring items from transaction stream

The work presents a simple but novel scheme to determine the recent frequent cooccurring items from transaction streams. It enables to follow the current trend, pattern of occurrence of the items and also help in recommendation. Proposed methodology works on a sequence of transaction forming a time-window. Based on the items present in a transaction, item-sets are formed and encoded. An in-memory structure maintains the frequency of the item-sets. Frequency is time-decayed count that prioritises the recent occurrences. To limit the size of in-memory structure, a pruning operation is carried out to remove the less frequent item-sets. Threshold for pruning is chosen dynamically. Impacts of decaying parameter and window size are also analysed. Experiment has been carried out on two datasets and performance has been compared with two recent works. Proposed methodology provides a better accuracy. Low time and space complexity make it suitable for stream analytic.