Temporal-Aware Learned Indexes for Time-Evolving Data

Current learned indexes assume static data distributions, failing in time-evolving scenarios (e.g., IoT sensor streams). This idea introduces temporal-aware models that predict key positions using both value and time features. For example, a hybrid model could combine a linear regression on key values with an LSTM on insertion timestamps, anticipating distribution shifts. This addresses Guo et al.'s (2024) Chameleon, which handles local skew but not temporal trends. Unlike SLIN's string-slicing (Song et al. 2024), we focus on temporal locality. By training on historical evolution patterns (e.g., seasonal trends), the index can proactively adjust its predictions. Experiments could show 3× lower retraining frequency than ALEX in time-series workloads, opening new applications for learned indexes in streaming analytics.

References:

1. SLIN: A CPU-efficient, Hybrid Tree and Learned Index for String Data. Yuanyuan Song, Miao Cai, Baoliu Ye, Guo Cheng (2024). International Conference on Compute and Data Analysis.
2. Chameleon: Towards Update-Efficient Learned Indexing for Locally Skewed Data. Na Guo, Yaqi Wang, Wenli Sun, Yu Gu, Jianzhong Qi, Zhenghao Liu, Xiufeng Xia, Ge Yu (2024). IEEE International Conference on Data Engineering.