Learning-based Sketches for Frequency Estimation in Data Streams without Ground Truth

The frequency or volume estimation of unending data streams is a concern in many domains, starting with telecommunications but spreading to social networks, finance, and learning-augmented streaming algorithms [10-15] is receiving website engine. In network fields, for example, professionals significant attention due to the powerful potential of machine want to keep track of the activity frequency to identify overall learning (ML) to relieve or eliminate the binding of data network health and potential anomalies or changes in behavior, characteristics and the sketch design. Their typical workflow which, however, is often challenging because the amount of involves training a heavy hitter oracle, which receives a key information may be too large to store in an embedded device and returns a prediction of whether it will be heavy or not, then or to keep conveniently in fast storage [1]. As a consequence, inserts the most frequent keys into unique buckets and applies sketch, which is a set of counters or bitmaps associated with a sketch to the remaining keys. Although filtering heavy items hash functions, and a set of simple operations that record has been proven to improve the overall sketch performance on approximate information [2], has grown in popularity in the heavy-tailed distribution [4, 10], these offline and supervised context of high-velocity data streams and limited computational methods could hardly work in real-world applications.