Learned index for non-key queries

Learned indexes have attracted a lot of interest lately due to their superior performance over conventional indexes. When there is a lot of data traffic, the learned index efficiently addresses the issue of the standard index's large memory usage. In this paper, we concentrate on a well-known learned index, the recursive model index (RMI). Since the machine learning model is unbiased while calculating, when there are too many non-key queried, the model will calculate the position of the key as if it were positive key, which wastes a lot of time on unnecessary calculations. To deal with this condition, we propose a hierarchical learned index structure based on Bloom filter named HBFdex. HBFdex can effectively prune non-keys, which means most non-key return in layer of BF before they get to machine learning model. By lowering the number of layers traversed by non-key and the time spent looking for non-key within the error bound that is provided by machine learning model, HBFdex decreases the average query time of learned index. We compare HBFdex with B-Tree and RMI, and the results prove that our new structure optimizes the performance of RMI in the case of non-key queries.