In-stream frequent itemset mining with output proportional memory footprint

We propose an online partial counting algorithm based on statistical inference that approximates itemset frequencies from data streams. The space complexity of our algorithm is proportional to the number of frequent itemsets in the stream at any time. Furthermore, the longer an itemset is frequent the closer is the approximation to its frequency, implying that the results become more precise as the stream evolves. We empirically compare our approach in terms of correctness and memory footprint to CARMA and Lossy Counting. Though our algorithm outperforms only CARMA in correctness, it requires much less space than both of these algorithms providing an alternative to Lossy Counting when the memory available is limited.