Learned sketches for frequency estimation

The Count-Min sketch and its variations are widely used to solve the frequency estimation problem due to its sub-linear space cost. However, the collisions between high-frequency and low-frequency items introduce a significant estimation error. In this paper, we propose two learned sketches called the Learned Count-Min sketch and Learned Augmented sketch. We combine the machine learning methods with the traditional Count-Min sketch and Augmented sketch to improve the performance. We used a regression model trained from historical data to predict the frequencies and separate the high-frequency items and low-frequency items. The experimental results indicated that our learned sketches outperform the traditional Count-Min sketch and Augmented sketch. The learned sketches can provide a more accurate estimation with a more compact synopsis size.