Mining Frequent Itemsets Over Arbitrary Time Intervals in Data Streams

Mining frequent itemsets over a stream of transactions presents di cult new challenges over traditional mining in static transaction databases. Stream transactions can only be looked at once and streams have a much richer frequent itemset structure due to their inherent temporal nature. We examine a novel data structure, an FP-stream, for maintaining information about itemset frequency histories. At any time, requests for itemsets frequent over user-de ned time intervals can be serviced by scanning the maintained FP-stream producing an approximate answer with error guaranteed to be no worse than a user-speci ed frequency and temporal threshold. We develop an algorithm for constructing and updating an FP-stream structure and present experiments illustrating the time and space required for maintenance.