Decentralised Preference Discovery Protocols

• Decentralised Preference Discovery Protocols are distributed algorithms that aggregate individual preferences without central coordination, using iterative local exchanges to reveal consensus and local structures.
• Epidemic, sampling, and cryptographic methods like gossip, TraceRank, and TRW+PSI underpin these protocols, achieving empirical convergence and robust performance against adversarial manipulation.
• Practical applications include peer-to-peer recommendation, decentralized marketplaces, privacy-preserving social networking, and matching markets with verifiable Sybil resistance.

Decentralised preference discovery protocols are distributed algorithms and frameworks designed to elicit, aggregate, and reveal individual or collective preferences in the absence of a trusted, central coordinator. These protocols are foundational in distributed systems, peer-to-peer recommendation, computational social choice, privacy-preserving friend/community detection, Sybil-resistant reputation systems, and decentralized matching markets. Their primary objective is to achieve robust, scalable, and often privacy-preserving preference aggregation—despite challenges such as asynchrony, partial information, censorship, and adversarial behavior.

1. Formal Models and Foundational Challenges

Decentralised preference discovery (DPD) involves a set of autonomous agents (nodes, peers, users) possessing private, subjective preferences—potentially as rankings, ratings, or utilities—over a set of alternatives or counterparts. The DPD objective is to uncover either a consensus outcome (e.g., winner, ranking, or matching) or relevant local/global preference structures, without centralized authority or full information visibility (Kotsialou, 20 Dec 2025).

Prominent challenge dimensions include:

• Information locality: No agent sees the global profile; sampling or bandit feedback dominates.
• Asynchrony: Agents act without global synchronization, updating out-of-phase.
• Censorship and fault tolerance: Protocols must ensure convergence despite node failures and message loss (Kotsialou, 20 Dec 2025).
• Adversarial manipulation: Defenses against Sybil attacks and privacy breaches are central (Shi et al., 31 Oct 2025, Hu et al., 2014).
• Learning and discovery: Agents may not know their own preferences a priori, necessitating integrated online learning mechanisms (Shah et al., 2024, Grenet et al., 2021).

2. Gossip, Epidemic, and Sampling-Based Protocols

A core design from DPD literature is the use of epidemic or gossip-based algorithms, where agents iteratively exchange information with random peers to foster a global consensus or high-quality neighborhood formation.

• Swarmix Protocol: In peer-to-peer recommenders, each peer maintains a cache of like-minded neighbors' profiles, updated via symmetric push–pull gossip. Each round, a peer merges its cache with that of a random neighbor, retaining the $k$ most similar profiles (cosine similarity) (0812.4460). This mechanism guarantees non-decreasing local utility, preservation of local fixed points, and empirical convergence to ≈90% centralized quality in $O(\log{|V|})$ cycles. The emergent overlay graph exhibits small-world properties and strong resilience to churn and partial network failure.
• Snowveil Framework: In social choice, agents repeatedly sample $k\ll n$ random peers' current choices, aggregate via a rule (e.g., Constrained Hybrid Borda), and probabilistically "lock" to a candidate when democratic support is robust (Kotsialou, 20 Dec 2025). The protocol provably converges to a single-winner or committee outcome in $O(n)$ steps under minimal sampling. Its modularity permits plug-in of any aggregation rule with positive responsiveness.

3. Sybil-Resistant Decentralised Reputation and Service Discovery

A salient category applies decentralized preference discovery to service, API, or agent-market selection, integrating strong Sybil-resistance.

• TraceRank Algorithm: Construct a payment flow graph with addresses as nodes and value/time-decayed payment edges (Shi et al., 31 Oct 2025). Reputable agents inject non-negative seed reputation ($s_i$), and steady-state reputation scores ($r$) are computed iteratively:

$r = s + \alpha W^\top r, \quad \alpha\in(0,1)$

Spam services that only receive endorsements (payments) from low- or zero-reputation addresses cannot inflate their ranks, as zero-seed nodes propagate no mass. Empirically, TraceRank achieves high precision, recall, and nDCG even under large-scale Sybil attacks—outperforming both transaction volume and semantic-only rankings. Hybridizing with embedding-based semantic search yields the best end-to-end results and spam robustness.

Approach	Sybil Resistance	Centralization	Empirical Precision@10
Volume Only	Low	None	0.45
Semantic Only	Moderate	None	0.62
TraceRank	High	None	0.75
Hybrid	High	None	0.82

Spam-induced degradation: TraceRank <5%, Hybrid <3% (Shi et al., 31 Oct 2025).

4. Privacy-Preserving Decentralised Community and Friend Discovery

Privacy-preserving protocols for preference/community discovery utilize cryptographic primitives and local random-walks to limit information leakage.

• TRW+PSI Protocol (Hu et al., 2014): Nodes in a social graph perform truncated random walks (TRW) to generate Walker-ID sets reflecting community structure. Nodes test for shared community membership by executing Private Set Intersection-Threshold (PSI-TH) protocols, which only reveal significant overlap (above threshold $T$) without exposing raw sets. Accuracy and privacy trade-offs are analytically tractable, with false negative/positive rates and privacy-advantage scaling with parameters $(W,L,T)$ and community size. The method generalizes to decentralized preference similarity retrieval for social, recommender, or mobility data.

5. Decentralized Preference Discovery in Matching Markets

Modern matching platforms often lack centralized control or agent transparency, requiring fully decentralized, learning-centric protocols.

• Bandit-Feedback Stable Matching (Shah et al., 2024): Proposers learn their unknown preference orderings over acceptors via repeated proposals and localized utility feedback. An uncoupled, asynchronous process—driven by a three-mood (Content/Discontent/Watchful) online state machine and low-rate experimentation—guarantees stochastic convergence to the proposer-optimal stable match (POSM) under only local, bandit feedback. All processes are decentralized, independent, and only require agents to monitor their own past utilities. The formal result relies on regular perturbation theory and resistance tree analysis, establishing that the POSM is uniquely stochastically stable.
• Preference Discovery in Admissions Matching (Grenet et al., 2021): In hybrid decentralized/centralized clearinghouse mechanisms, early decentralized offer rounds enable students to learn costly private match-values over time, with final centralized deferred acceptance rounds guaranteeing stability and strategy-proofness. Empirically, early offers are more likely to be accepted, consistent with costly preference discovery and decreasing acceptance thresholds over time in a dynamic search model.

6. Theoretical Guarantees, Scalability, and Robustness

Across protocols, analysis typically addresses scalability, convergence, accuracy, privacy, and robustness:

• Convergence: Gossip-based protocols (e.g., Snowveil) achieve almost-sure convergence in $O(n)$ rounds to consensus or $k$-winner outcomes, under minimal axiomatic conditions such as determinism and positive responsiveness (Kotsialou, 20 Dec 2025).
• Scalability: Epidemic and random-sampling methods incur per-peer communication costs independent or logarithmic in network size (0812.4460, Kotsialou, 20 Dec 2025).
• Privacy/Sybil-resistance: Protocols such as TraceRank and TRW+PSI provide rigorous analytic bounds on resistance to Sybil attacks or privacy breaching adversaries, facilitated by seed mass constraints and cryptographically-limited information exposure (Shi et al., 31 Oct 2025, Hu et al., 2014).
• Empirical results: Protocols are validated on large datasets (e.g., MovieLens, x402 payment graphs), demonstrating near-centralized quality, minimal communication requirements, and robustness to network churn and adversarial activity (0812.4460, Shi et al., 31 Oct 2025, Kotsialou, 20 Dec 2025).
• Multi-winner and dynamic extension: Frameworks extend to sequential or committee selection by iterating the base protocol, ensuring modular applicability (Kotsialou, 20 Dec 2025).

7. Practical Applications and Design Considerations

Applications span recommender systems, decentralized marketplaces, privacy-respecting social networks, agent economies, and distributed voting or committee selection:

• Peer-to-peer recommendation: Epidemic protocols autonomously form high-quality, similarity-based recommendation neighborhoods without central orchestration (0812.4460).
• Service and API discovery: Sybil-resistant reputation propagation reveals high-trust services in blockchain and open economic networks (Shi et al., 31 Oct 2025).
• Matching markets and admissions: Online learning and dynamic multi-offer mechanisms facilitate efficient, strategy-proof matching with endogenous preference discovery (Shah et al., 2024, Grenet et al., 2021).
• Secure social networking: TRW+PSI constructs enable discovery of friends or communities with quantifiable privacy guarantees (Hu et al., 2014).
• Social choice and governance: Modular gossip frameworks enable censorship-resistant, scalable voting, supporting a wide class of aggregation rules (Kotsialou, 20 Dec 2025).

Parameter selection, governance of seed assignment, privacy-communication trade-offs, and handling of dynamic or adversarial environments are critical in protocol design. Many frameworks are flexible, supporting the integration of new aggregation rules, cryptographic primitives, or network substrates. A key general insight is that decentralized, iterative local operations—underpinned by monotonicity, stochastic stability, or axiomatic social choice theory—can reliably uncover global preference structure in unstructured, adversarial, and information-sparse environments.