Continuous Fairness On Data Streams

TL;DR

This work initiates the study of block-level group fairness on data streams, enforcing fairness inside each sliding window by partitioning the window into $k$ blocks of size $s$ and enforcing target proportions via $f(p)$. It combines a sketch-based continuous fairness monitor (Monitor-BFair) with an optimal continuous stream reordering mechanism (BFair-ReOrder) that can use landmark items $ ext{X}$ to extend the stream and maximize the number of unique fair blocks. The framework provides formal definitions, theoretical guarantees (including optimality of isomorphic/extended isomorphic streams and linear-space, near-linear-time monitoring), and empirical validation across four real-world datasets, achieving millisecond latency and throughput up to roughly $3 imes 10^{4}$ queries/second. The results demonstrate substantial gains in block-level fairness compared to window-level baselines and show practical viability for high-throughput streaming systems, with clear guidance on parameter choices (window size $|W|$, block size $s$, landmark size $| ext{X}|$). Overall, the paper delivers a principled, scalable approach to continuous fairness enforcement in streaming environments with real-time guarantees and demonstrated effectiveness.

Abstract

We study the problem of enforcing continuous group fairness over windows in data streams. We propose a novel fairness model that ensures group fairness at a finer granularity level (referred to as block) within each sliding window. This formulation is particularly useful when the window size is large, making it desirable to enforce fairness at a finer granularity. Within this framework, we address two key challenges: efficiently monitoring whether each sliding window satisfies block-level group fairness, and reordering the current window as effectively as possible when fairness is violated. To enable real-time monitoring, we design sketch-based data structures that maintain attribute distributions with minimal overhead. We also develop optimal, efficient algorithms for the reordering task, supported by rigorous theoretical guarantees. Our evaluation on four real-world streaming scenarios demonstrates the practical effectiveness of our approach. We achieve millisecond-level processing and a throughput of approximately 30,000 queries per second on average, depending on system parameters. The stream reordering algorithm improves block-level group fairness by up to 95% in certain cases, and by 50-60% on average across datasets. A qualitative study further highlights the advantages of block-level fairness compared to window-level fairness.

Figures (25)

• Figure 1: The proposed framework consists of two tightly coupled components. For the first ever window, it constructs a compact fairness sketch that summarizes block-level counts; for each subsequent window, this sketch is incrementally updated as the window slides. This sketch is used to perform continuous fairness monitoring, enabling efficient verification of block-level group fairness constraints within each window. When a window violates fairness, the framework first attempts to perform continuous stream reordering using only the items in the current window. If no feasible reordering exists, it reads an additional set of landmark items and jointly reorders the window and landmarks. The reordering objective is to maximize the number of unique fair blocks across the current window and the additional windows induced by the landmark items.
• Figure 2: A sliding window of $15$ items (contents) with $3$ blocks, consisting of contents from three ethnic groups $[\emph{C}$ (Caucasian), $\emph{A}$ (Asian), $\emph{H}$ (Hispanic)].
• Figure 3: Five landmark items $\mathcal{X}$
• Figure 4: Sketch of blocks $B_1$ and $B_2$ in Figure \ref{['fig:windowblock']}
• Figure 5: Sketch maintenance with one item slide

Theorems & Definitions (10)

• definition 1: Forward Sketch
• theorem 1
• lemma 1
• lemma 2
• lemma 3
• lemma 4
• definition 2: Isomorphic Blocks and Isomorphic Block Count
• definition 3: Isomorphic Stream and Extended Isomorphic Stream
• lemma 5: Optimality of Isomorphic and Extended Isomorphic Streams
• theorem 2: Algorithm BFair-ReOrder is optimal