A Sampling Method of Finding Top-k Frequent Items on Timestamp-Based Stream

Abstract

Data streams with high volume and complicated items become more and more common, and typical algorithms of finding top-k frequent items on streams, such as counter-based algorithms and sketch algorithms, are gradually not keeping up with efficiency requirements. Our paper focuses on finding top-k frequent items on timestamp-based complicated streams, and proposes an approximate solution by sampling. Specifically, we design a multi-treap parallel priority algorithm to maintain uniform sample on timestamp-based sliding windows. The top-k answers are approximated through processing on samples. We also theoretically analyze the relationship between item accuracy and sample size. Through experimental analysis on real data, our method provides flexible sample size to satisfy different accuracy requirements and ensure a good running efficiency.