[Paper Review] Querying with Conflicts of Interest

Nischal Aryal, Arash Termehchy+1 more|arXiv (Cornell University)|Mar 5, 2026

Data Quality and Management0 citations

TL;DR

The paper proposes a formal framework for querying when data sources have incentives to bias results, develops efficient algorithms to detect influential interactions and untrustworthy results, and to reformulate queries to improve information usefulness, with empirical validation on real datasets.

ABSTRACT

Conflicts of interest often arise between data sources and their users regarding how the users' information needs should be interpreted by the data source. For example, an online product search might be biased towards presenting certain products higher than in its list of results to improve its revenue, which may not follow the user's desired ranking expressed in their query. The research community has proposed schemes for data systems to implement to ensure unbiased results. However, data systems and services usually have little or no incentive to implement these measures, e.g., these biases often increase their profits. In this paper, we propose a novel formal framework for querying in settings where the data source has incentives to return biased answers intentionally due to the conflict of interest between the user and the data source. We propose efficient algorithms to detect whether it is possible for users to extract relevant information from biased data sources. We propose methods to detect biased information in the results of a query efficiently. We also propose algorithms to reformulate input queries to increase the amount of relevant information in the returned results over biased data sources. Using experiments on real-world datasets, we show that our algorithms are efficient and return relevant information over large data.

Motivation & Objective

• Model how user intents and data source biases interact when returning ranked query results with conflicts of interest.
• Define stable (Bayesian equilibrium) states of user and data source reasoning in this setting.
• Develop algorithms to detect influential interactions and untrustworthy results.
• Prove complexity results for finding queries that maximize user information gain.
• Empirically validate scalability and effectiveness on real-world datasets.

Proposed method

• Model the user and data source as agents with utilities U^r and U^s over interpretations beta of a query q.
• Use Bayesian equilibrium concepts to characterize stable interaction states (tau, beta).
• Provide necessary and sufficient conditions (Theorem 3.1) for influential interactions.
• Introduce bias functions b(e) and additive utility form U^s and U^r to analyze bias influence (Definition 3.3, Proposition 3.5).
• Develop efficient tests (Theorem 3.6 and Corollary 3.7) to detect non-influential interactions under convex/quadratic-like loss L.
• Define trustworthiness of results (Definition 4.1) and propose methods to detect untrustworthy tuples.

(a) Amazon

Experimental results

Research questions

• RQ1 Under what conditions do user and data source interactions converge to a stable influential equilibrium?
• RQ2 When can a data source’s bias render the user unable to influence the returned results (non-influential equilibrium)?
• RQ3 How can one detect untrustworthy results and quantify trustworthy information in biased interpretations?
• RQ4 What is the computational complexity of finding a query that yields the maximum information about the user’s intent, and can efficient solutions be found for broad utility classes?
• RQ5 How scalable are the proposed algorithms on large real-world datasets?

Key findings

• An equilibrium is influential if there exist set-equivalent intents tau and distinct interpretations beta, beta′ satisfying user and data source payoff conditions (Theorem 3.1).
• If the data source bias is extremely strong, there may be no influential equilibrium, meaning users cannot persuade the source to change its output (Theorem 3.6).
• Corollary 3.7 provides constant-time checks for non-influential interactions under certain convex, non-decreasing loss functions.
• Bias functions and supermodular utilities model partial common interest and allow efficient analysis of influence and trustworthy results (Propositions 3.5, 3.6).
• Definition 4.1 formalizes untrustworthy results, enabling detection of misranked or omitted tuples in beta(I).
• Empirical studies show the framework and algorithms scale to large data and yield relevant information over real-world datasets (Section 6).

(b) Pricerunner