Harm-Aware Ranking Accuracy

• Harm-aware ranking accuracy is a framework that combines classical relevance measures with harm-centric metrics to explicitly address social and group-level biases.
• Core methodologies include adversarial debiasing, group-conditioned recovery, and regret-aware optimization to balance ranking performance with fairness constraints.
• Applications span news recommendation, social media moderation, and multi-stakeholder ranking, highlighting trade-offs managed through precise hyperparameter tuning.

Harm-aware ranking accuracy refers to the suite of methodologies and metrics that explicitly quantify and optimize a recommender or ranking system’s utility under constraints that address user harm, exposure fairness, or the propagation of undesirable content. While traditional ranking accuracy concerns itself with matching user intent or maximizing engagement, harm-aware ranking accuracy formalizes the intentional mitigation of social, individual, or group-level harms—by surfacing non-harmful content, minimizing group-based bias, or ensuring equitable error across subpopulations.

1. Formal Definitions and Foundations

Harm-aware ranking accuracy incorporates both conventional ranking metrics and additional harm/fairness-centered measures. Classical measures such as AUC and nDCG@K evaluate the relevance or quality of rankings, but are agnostic to the social costs of misranking protected classes or harmful content. Harm-aware accuracy extends the evaluation to include:

• Group-conditioned Kemeny error: Measures rank disagreement within or across subgroups, defining for any group $G$ the normalized error

$D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$

where $t$ is latent ground-truth skill and $r$ the recovered rank (Ahnert et al., 2024).

• Demographic parity and equality of opportunity gaps: For recommendation or ranking tasks, these gaps quantify disparities in recommendation rates or true positive rates between protected groups:

$$\begin{align*} &\text{Demographic Parity Gap:}\;\; |P(\mathrm{recommend}|Z=0) - P(\mathrm{recommend}|Z=1)| \ &\text{Equality of Opportunity Gap:}\;\; |\mathrm{TPR}_{Z=0} - \mathrm{TPR}_{Z=1}| \end{align*}$$

(Wu et al., 2022).

• Harm-centric metrics: Such as Top-Pref-$k$ (TP$_k$), Per-Pref-$k$ (PP$_k$), and Exponentially Weighted Normalization (EWN), all of which prioritize non-harmful content in ranked outputs (Oak et al., 23 Jan 2025).
• Regret-normalized perceived satisfaction: Drawing from regret theory, user-perceived utility penalizes deviation from the ideal list:

$$Z(\pi_n) = q_{\pi_n} + 1 - \exp(-\delta[q_{\pi_n} - q_{\pi_n}^*])$$

where $D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$0 is user $D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$1’s observed quality and $D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$2 is the ideal (Ye et al., 20 Apr 2025).

These analytic definitions enable harm-aware ranking systems to optimize utility while constraining or directly minimizing instances of user or group harm.

2. Core Methodologies for Harm-Aware Ranking

Achieving high accuracy under harm constraints leverages a variety of algorithmic techniques:

• Adversarial debiasing in representation: As exemplified in FairRank (Wu et al., 2022), candidate-aware models employ dual adversarial paths (candidate-aware and candidate-invariant user embeddings), gradient reversal, and KL alignment regularizers to systematically strip sensitive-attribute signals from user and item representations, while minimizing loss in ranking performance.
• Group-conditioned recovery from biased pairwise data: Fairness-Aware PageRank enforces group-balanced teleportation in rank propagation, preventing overexposure or underexposure of protected groups (Ahnert et al., 2024). GNNRank post-processed with FA*IR further constrains exposure or proportion constraints with minimal swaps.
• Regret-aware multi-stakeholder optimization: The BankFair+ framework reformulates list re-ranking as a regret-aware fuzzy programming problem, introducing non-linear regret-rejoice utilities to ensure individual user fairness alongside provider fairness (exposure proportional to merit) (Ye et al., 20 Apr 2025).
• Harm-based content re-ranking: LLM-powered re-ranking, through zero-shot or few-shot preference comparisons, demotes harmful content in output sequences, directly optimizing harm-centric metrics without labeled training data (Oak et al., 23 Jan 2025).

Algorithmic frameworks supporting harm-aware ranking are ably summarized in Table 1.

Method	Harm/Accuracy Focus	Key Mechanism
FairRank	Group fairness, causal harm	Adversarial, KL alignment
Fairness-Aware PageRank	Group exposure and error parity	Group-balanced teleportation
GNNRank + FA*IR	Error minimization, exposure parity	Post-processing re-ranking
BankFair+	Individual and provider fairness	Regret-based fuzzy opt.
LLM Harm Re-ranking	Content safety (early exposure)	Zero-/few-shot pairwise comp.

3. Harm-Aware Ranking Metrics

Three classes of metrics dominate recent evaluation frameworks:

1. Classical Relevance Metrics:
   • AUC, nDCG@K (Wu et al., 2022, Ye et al., 20 Apr 2025).
2. Fairness and Harm-Aware Metrics:
   • Group-conditioned Kemeny error $D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$3 and overall $D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$4 (Ahnert et al., 2024).
   • Exposure difference $D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$5 quantifying representation at early ranks.
   • Demographic parity and equality of opportunity gaps (Wu et al., 2022).
3. Content-Harm Metrics (Oak et al., 23 Jan 2025):
   • TP$D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$6: Fraction of non-harmful items among top $D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$7.
   • PP$D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$8: Required prefix length to encounter $D^G_t(r) := \sqrt{\frac{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2\,\mathbb{I}[(t_i-t_j)(r_i-r_j)>0]}{\sum_{\substack{i<j\i\in N,\;j\in G}}(t_i-t_j)^2}} \in [0,1]$9 harmful items.
   • EWN: Positionally weighted, sequence-normalized harmlessness.

Metrics such as Min–Max Ratio (MMR@K) and user NDCG variance (Ye et al., 20 Apr 2025) directly capture the equity of accuracy across user subpopulations.

4. Empirical Trade-Offs and Pareto Frontiers

Empirical studies across domains consistently reveal trade-offs between traditional accuracy and harm-sensitive objectives:

• FairRank (Wu et al., 2022): Introduction of adversarial and KL losses yields substantial drops in gender-inference accuracy (e.g., Acc@10 from 56.7% to 52.4% on FairNews), with only a minor decrease in AUC (≤0.4 points). Ablations confirm the necessity of candidate-invariant adversarial branches and the efficacy of KL alignment in augmenting fairness.
• Fair Pairs (Ahnert et al., 2024): Bias injections yield measurable gaps ($t$0) in group-conditioned error and exposure, which Fairness-Aware PageRank nearly nullifies (e.g., $t$1 from ~0.05 to ~0.01), at small cost to total error. GNNRank with FA*IR minimizes overall error, but reintroduces some group disparity.
• BankFair+ (Ye et al., 20 Apr 2025): Regret-aware re-ranking increases individual fairness (MMR rises from 0.49 to 0.71 on KuaiRand-1K) and mean NDCG (+0.20 at fixed provider fairness), while allowing explicit parameter control over the balance with provider-side objectives (ESP, Gini).
• LLM Harm Re-ranking (Oak et al., 23 Jan 2025): Across datasets and configurations, LLM-based approaches consistently improve harm metrics (up to +0.25 in TP$t$2, +0.13 in PP$t$3), with marginal impact on aggregate relevance. Notably, robustness to rising harm rates is significantly greater for harm-aware methods.

A key operational principle is the tuning of trade-off hyperparameters (e.g., $t$4 in FairRank, $t$5 in BankFair+) to reach a desired point on the accuracy–harm Pareto frontier, optimizing for maximal utility at an acceptable harm or fairness constraint.

5. Algorithmic and Practical Insights

Harm-aware ranking accuracy frameworks share several practical properties and theoretical implications:

• Monotonicity and compositionality: Many harm-aware metrics (TP$t$6, PP$t$7, EWN) increase monotonically as harmful items are demoted. Group-conditioned error and exposure differences similarly decrease as harm-mitigating mechanisms are applied (Ahnert et al., 2024, Oak et al., 23 Jan 2025).
• Low additional computational burden: Frameworks such as FairRank and BankFair+ impose moderate cost increases, primarily in adversarial or fuzzy-programming modules, but retain efficiency relative to baseline ranking models (Wu et al., 2022, Ye et al., 20 Apr 2025).
• Generalizability: Harm-aware objectives are extensible to multiple harm/fairness definitions. Adversarial losses, regret-penalized utilities, and group-informed teleportation can target distinct or multiple protected groups without major architectural changes (Wu et al., 2022, Ahnert et al., 2024).
• Evaluation interpretability: Harm-based metrics provide interpretable, threshold-free decision rules (e.g., “maximize TP$t$8 above 0.85 while minimizing AUC drop”), enabling actionable moderation policies or fair ranking deployments (Oak et al., 23 Jan 2025).

6. Applications and Limitations

Applications of harm-aware ranking accuracy span:

• News recommendation: Reducing sensitive-attribute leakage and group overexposure via dual adversarial debiasing and KL regularization (Wu et al., 2022).
• Social media moderation: Demoting harmful content in feed ranking with LLM-based zero- and few-shot re-ranking, evaluated under explicit harm exposure metrics (Oak et al., 23 Jan 2025).
• Pairwise comparison ranking: Fairness-aware recovery from human-annotated pairwise data, correcting for group biases in both simulated and empirical datasets (Ahnert et al., 2024).
• Multi-stakeholder recommendation: Regret-aware list re-ranking ensuring both individual user fairness and provider exposure proportionality (Ye et al., 20 Apr 2025).

Principal limitations include:

• Sensitivity to latent bias structure: On real-world data lacking perfect group balance, harm-aware methods (e.g., Fairness-Aware PageRank) may increase overall error (Ahnert et al., 2024).
• Trade-off parameter selection: Hyperparameter tuning is dataset- and context-specific, with no universally optimal setting (Wu et al., 2022, Ye et al., 20 Apr 2025).
• Scarcity of theoretical guarantees for optimality under complex harm and fairness constraints; practice-driven annular evaluation is standard.

7. Outlook and Extensions

Harm-aware ranking accuracy frameworks redefine evaluation and optimization in ranking systems by embedding harm, fairness, and exposure criteria as co-equal with relevance. Ongoing research explores:

• Extensions to intersectional and dynamic notions of harm, with multi-attribute debiasing (Wu et al., 2022, Ahnert et al., 2024).
• Automated, preference-informed trade-off selection via meta-learning or policy optimization.
• Robustness and efficacy in the face of adversarial attacks, concept drift, and evolving definitions of content harm (Oak et al., 23 Jan 2025).

A plausible implication is that, as harm-aware objectives mature and deployment scaling increases, ranking system audits and regulatory standards will increasingly demand explicit validation against these measures, driving wider adoption in real-world platforms.

---

References:

(Wu et al., 2022): "FairRank: Fairness-aware Single-tower Ranking Framework for News Recommendation" (Ahnert et al., 2024): "Fair Pairs: Fairness-Aware Ranking Recovery from Pairwise Comparisons" (Oak et al., 23 Jan 2025): "Re-ranking Using LLMs for Mitigating Exposure to Harmful Content on Social Media Platforms" (Ye et al., 20 Apr 2025): "Regret-aware Re-ranking for Guaranteeing Two-sided Fairness and Accuracy in Recommender Systems"