Guaranteeing Accuracy and Fairness under Fluctuating User Traffic: A Bankruptcy-Inspired Re-ranking Approach

TL;DR

BankFair addresses the challenge of guaranteeing both short-term user accuracy and long-term provider fairness under fluctuating user traffic. It models exposure allocation as a sequential bankruptcy problem and applies the $Talmud$ rule to distribute fairness across time, coupled with an online re-ranking algorithm that enforces $a(n)\ge \phi$ and $\sum_{n=1}^N \bm{E}_{p,n} \ge \bm{m}_p$. On two real datasets, BankFair achieves higher $NDCG@K$ and lower $Vio@K$ while achieving $ESP=1$ (100%) for all providers, demonstrating robustness to traffic fluctuations under the joint constraints. This work advances practical two-sided fairness in dynamic recommender systems and provides a scalable framework for industry deployment.

Abstract

Out of sustainable and economical considerations, two-sided recommendation platforms must satisfy the needs of both users and providers. Previous studies often show that the two sides' needs show different urgency: providers need a relatively long-term exposure demand while users want more short-term and accurate service. However, our empirical study reveals that previous methods for trading off fairness-accuracy often fail to guarantee long-term fairness and short-term accuracy simultaneously in real applications of fluctuating user traffic. Especially, when user traffic is low, the user experience often drops a lot. Our theoretical analysis also confirms that user traffic is a key factor in such a trade-off problem. How to guarantee accuracy and fairness under fluctuating user traffic remains a problem. Inspired by the bankruptcy problem in economics, we propose a novel fairness-aware re-ranking approach named BankFair. Intuitively, BankFair employs the Talmud rule to leverage periods of abundant user traffic to offset periods of user traffic scarcity, ensuring consistent user service at every period while upholding long-term fairness. Specifically, BankFair consists of two modules: (1) employing the Talmud rule to determine the required fairness degree under varying periods of user traffic; and (2) conducting an online re-ranking algorithm based on the fairness degree determined by the Talmud rule. Experiments on two real-world recommendation datasets show that BankFair outperforms all baselines regarding accuracy and provider fairness.

Figures (7)

• Figure 1: Illustrative experiments based on KuaiRand dataset. (a) Lower user traffic leads to more accuracy loss; (b) Applying distinct fairness strategies based on different traffic levels.
• Figure 2: Illustration of algorithm updates over time. The shadow area represents the user traffic of each interval. The minimum exposure needs to be guaranteed at the end of the $N$-th interval.
• Figure 3: A toy example with two providers to illustrate the optimization process at interval $i$ of Equation \ref{['eq:unfied_opt']}. Provider 1 (x-axis) requires a minimum exposure guarantee of $\bm{M}_{1,i}=4$ and provider 2 (y-axis) has no requirement ($\bm{M}_{2,i}=0$). The green area is the feasible region constructed by the fairness constraint. (a) Suppose 3 users arrive at interval $i$ and each is recommended 5 items ($3\times 5 = 15$ total exposures). The grey point denotes the accuracy solution $(\bar{\bm{E}}_{1,i},\bar{\bm{E}}_{2,i})$ and the orange point denotes the provider-fair solution $(\bm{E}^{\text{fair}}_{1,i},\bm{E}^{\text{fair}}_{2,i})$, verifying Theorem 1; (b) Suppose the number of users reduced to 2 at interval $i+1$ ($2\times 5 = 10$ total exposures). The red point denotes the provider-fair solution, verifying Theorem 2.
• Figure 4: Workflow of the proposed BankFair model
• Figure 5: Pareto frontier of two different datasets with different top-$K$ ranking. Y-axis shows ESP@K metric, while X-axis shows NDCG@K metric and Vio@K metric. $\uparrow$ means higher values are better and $\downarrow$ favors lower values. According to the size of the dataset, for KuaiRand-1K, we set $\bm{m}_p = 1000$ and for Huawei-Video, we set $\bm{m}_p = 100$. For both datasets, we set $\phi=0.95$.

Theorems & Definitions (4)

• Theorem 1
• Theorem 2
• Theorem 3
• Lemma 1