Updatable Spatial Learned Index Based on Dimensionality Reduction

With the advent of the big data era, efficient data retrieval from massive datasets has become increasingly crucial. Traditional indexing structures, such as R-trees, offer conve-nience for data retrieval but often come with substantial storage requirements and higher re-insertion costs. Recent advancements in technologies, particularly machine learning, have introduced new optimization methods for index. These methods leverage predictive approaches to replace traditional traversal lookups. Experiments have demonstrated that such predictive methods can significantly reduce storage costs while maintaining retrieval efficiency comparable to traditional indexing structures. In light of this, our paper proposes a novel approach called USLI-DR (Up datable Spatial Learned Index based on Dimensional Reduction), which utilizes a machine learning technique in two-dimensional spatial data. Compared to the traditional R-tree index, USLI-DR not only improves retrieval and insertion efficiency but also reduces storage requirements by up to 98 %.