DWLBF: a deletable weighted learned Bloom Filter based on data temperature

The recently designed Learned Bloom Filter (LBF), which combines Machine Learning (ML) with the traditional BF to reduce memory footprint and improve overall performance, has garnered significant attention from researchers. Therefore, many LBFs are optimized based on the original LBF to improve the filter performance further. However, most LBFs do not consider the impact of data access frequency (called data temperature) on the overall performance of the model. In addition, efficient data deletion is another challenge for LBFs. Based on these issues, we proposed a novel Deletable Weighted Learned Bloom Filter (DWLBF) for large-scale membership query, which can reduce the False Positive Rate (FPR) and support deletion efficiently. In the learning model part, we designed a weighted learning model based on data temperature to effectively reduce the FPR. To ensure filter performance after data deletion, a perfect hash (PH) table based on data temperature was constructed before the learning model to store a part of the deleted data with a high access frequency. A series of experiments was conducted on malicious URLs, demonstrating that the proposed DWLBF method achieves the lowest FPR when querying without deleting data, and the FPR and query time consumption are greatly reduced after deleting data.