Learning-Augmented Moment Estimation on Time-Decay Models
Soham Nagawanshi, Shalini Panthangi, Chen Wang, David P. Woodruff, Samson Zhou
TL;DR
This paper introduces learning-augmented algorithms for time-decay models in streaming data, improving efficiency in weighted data scenarios like sliding windows, supported by theoretical analysis and empirical validation.
Contribution
It develops novel learning-augmented algorithms for fundamental streaming problems under time-decay models, addressing weighted data challenges with theoretical and empirical insights.
Findings
Improved space efficiency in weighted streaming models.
Effective algorithms for norm, moment, and frequency estimation.
Empirical results show practical efficiency on real and synthetic data.
Abstract
Motivated by the prevalence and success of machine learning, a line of recent work has studied learning-augmented algorithms in the streaming model. These results have shown that for natural and practical oracles implemented with machine learning models, we can obtain streaming algorithms with improved space efficiency that are otherwise provably impossible. On the other hand, our understanding is much more limited when items are weighted unequally, for example, in the sliding-window model, where older data must be expunged from the dataset, e.g., by privacy regulation laws. In this paper, we utilize an oracle for the heavy-hitters of datasets to give learning-augmented algorithms for a number of fundamental problems, such as norm/moment estimation, frequency estimation, cascaded norms, and rectangular moment estimation, in the time-decay setting. We complement our theoretical results…
Peer Reviews
Decision·ICLR 2026 Poster
1) Novel and timely problem formulation: First learning-augmented guarantees for time-decay streams. Captures privacy-driven data deletion (GDPR) and recency-weighted analytics, motivating practical relevance. 2) Theoretically strong results: General reduction shows any ($\alpha$, $\beta$)-smooth function enjoys a black-box sliding-window lift while preserving approximation and randomised guarantees; rectangle and cascaded norms benefit for free 3) Experimental validation on real data: CAIDA t
1) Limited empirical scope: Only l2 and l3 norms are evaluated; rectangle and cascaded-norm algorithms lack any implementation or micro-benchmark, leaving practical impact uncertain. 2) Expand empirical coverage: Include CPU-time and peak RAM tables for each dataset to confirm the claimed overhead. 3) Strengthen statistical reporting: Release full parameter files (hash seeds, repetition counts, bucket sizes) to facilitate exact reproduction.
- The extension of learning-augmented frequency moment estimation to general time-decay settings is novel and appealing. - The paper provides space complexity bounds for several common estimation tasks—including $F_p$ moments, rectangular moments, and cascaded norms—under two decay models, filling an existing theoretical gap. - The experiments convincingly show that the learning-augmented approach significantly improves estimation accuracy under the sliding-window model.
- The experimental evaluation does not include direct comparisons with other learning-augmented sliding-window algorithms, such as Shahout et al. (2024), and lacks empirical validation in more general time-decay environments. - The discussion on training suffix-compatible heavy-hitter oracles is insufficient. Although the PAC learning framework and Theorem 14 address the learnability of general heavy-hitter oracles, the paper does not examine specific training strategies or conditions that guar
Moment estimation is an important problem sketching problem and I find it well-motivated in the learning augmented-framework. It is well-studied in several past works. The sliding window model is also natural since it captures the idea that we may often be more interested in statistics of the most recent data. I like specifically about the paper, that they generalize the framework of Braverman and Ostrovsky [BOO7] to more general time-decay functions.
(1) I found that the paper lacked technical novelty. It heavily relies on the smooth histogram algorithm from Braverman and Ostrovsky [BOO7] which turns a classic sketching algorithm into an algorithm in the sliding window model with little overhead. The paper shows that the learning augmented algorithm from [JLL+20] fits the framework from [BO07] in a white-box fashion, they argue that the sliding-window algorithm from [BOO7] retains two properties captured in Proposition 1 and 2 even for the l
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsPrivacy-Preserving Technologies in Data · Data Stream Mining Techniques · Age of Information Optimization