Efficient Index Learning via Model Reuse and Fine-tuning

Learned indices using machine learning techniques have demonstrated potential as alternatives to traditional indices such as B-trees in both query time and memory. However, a well fitted learned index requires significant space consumption to train models and tune parameters. Furthermore, fast training methods—ones that train in one pass—may not learn the data distribution well. To consider both the fitness to data distribution and building efficiency, in this paper, we apply pre-trained models and fine-tuning to accelerate the building of learned indices by 30.4% and improve lookup efficiency by up to 24.4% on real datasets and 22.5% on skewed datasets.