PaperlayerView exact revision on arXiv

Source-linked AI summary

DeFiRank: Federated Pairwise Ranking for Governance in Decentralized Financial Networks

Zahra Batool, Michael Duchesne, Kaiwen Zhang, Ulrich Aivodji

arXiv:2610.12916v1cs.AI

TL;DR

Manual governance in decentralized financial networks relies on human review, while centralized training raises privacy and data-sovereignty concerns. DeFiRank uses federated pairwise ranking to learn provider rankings without sharing private transaction data. Its federated ranking is modestly below centralized NDCG@15 but matches Kendall τ stability, and the system produces promotion and demotion candidates.

  • Problem

    Manual governance depends on human review, while centralized model training requires institutions to share sensitive proprietary transaction records.

    Source: §I · p. 1

  • Method

    DeFiRank is a federated pairwise ranking framework that collaboratively learns provider rankings while preserving private transaction data.

    Source: §I · p. 2

  • Results

    NDCG@15 is ≈0.87 federated versus ≈0.92 centralized, while Kendall τ is ≈0.70 for both; the pipeline produced 648 promotion and 39 demotion candidates.

    Source: §I · p. 2

  • Takeaways & Limitations

    The framework learns network-wide behavioral patterns and produces governance decisions without exposing private transaction data.

    Source: §V · p. 8

Abstract

from arXiv · show

Existing governance approaches in decentralized financial networks rely on manual review processes conducted by human committees, which introduces subjective bias, creates a scalability bottleneck under high network volume, and fails to capture complex behavioural patterns and evolving network dynamics. To address these limitations, we propose \name, a federated learning-based pairwise ranking framework in which multiple institutions collaboratively train a ranking model without sharing their private transaction data. This collaborative learning enables the model to capture network-wide behavioural patterns that signify strong network participation and directly contribute to the network effect, including diversity of counterparties, usage of multiple applications, and frequency of transactions. The proposed system ranks providers based on the authenticity of their actions, promoting genuine contributors and demoting those exhibiting suspicious behaviour. We evaluate \name\ on a real-world blockchain transaction dataset. Results demonstrate that the federated approach achieves ranking performance modestly below the centralized baseline (NDCG@15: $\approx$0.87 vs.\ $\approx$0.92) while preserving data privacy and matching its ranking stability (Kendall $τ$: $\approx$0.70 for both regimes), producing reliable governance decisions, including provider promotion and demotion (648 and 39 unique providers, respectively), without requiring centralized access to sensitive transaction data.

I. INTRODUCTION

DeFiRank addresses manual governance and centralized-data constraints with privacy-preserving federated pairwise ranking, using network-wide behavioral signals to rank provider authenticity.

I. INTRODUCTION

  • Manual committee reviews introduce bias and struggle to adapt as network behavior and applications evolve, leaving governance decisions lagging network dynamics.The existing process promotes authentic contributors and demotes others using a human-crafted, single-party scoring function with predetermined weights.Source: §I · p. 1
  • Private provider interactions hide network-wide application-use and counterparty-diversity patterns from any single centralized observer, motivating federated learning across institutions.Federated learning can learn from localized transactions without hosts exposing raw data.Source: §I · p. 1
  • Pairwise ranking minimizes relative-ordering mistakes, framing governance as a comparison of which provider demonstrates greater authenticity.The ranking objective targets pairwise inversions and can identify contribution signals such as counterparty diversity, transaction volume, or application engagement.Source: §I · pp. 1–2
  • DeFiRank replaces manual epoch scoring with a federated pairwise ranker that learns strong-participation features while preserving providers’ transaction-data privacy.Named signals include counterparty diversity, use of multiple applications, and transaction frequency.Source: §I · p. 2
  • NDCG@15 is ≈0.87 federated versus ≈0.92 centralized, while Kendall τ is ≈0.70 for both; decisions included 648 promotion and 39 demotion candidates.Promotion consistency was 100% and demotion consistency 72% over 2+ consecutive days.Source: §I · p. 2

II. LITERATURE REVIEW

The review positions DeFiRank between privacy-preserving ranking and AI governance: prior work addresses each area, but not decentralized governance under structural heterogeneity and private data constraints.

A. Pairwise Ranking in Federated Learning

  • Federated pairwise-ranking methods target retrieval or recommendation tasks, with labels derived from clicks or ratings rather than governance decisions.The cited work includes online learning-to-rank, recommendation, browser URL suggestions, and compositional pairwise-risk optimization.Source: A. Pairwise Ranking in Federated… · p. 2

B. AI-Driven Corporate Governance

  • AI-driven governance work targets improved decision-making but predominantly assumes centralized data access, unlike decentralized financial settings.DeFiRank is positioned at the intersection of privacy-preserving ranking and governance objectives.Source: B. AI-Driven Corporate Governance · p. 2

III. PROPOSED METHOD

  • DeFiRank has three phases: federated temporal pairwise ranking across non-IID clients, local training with server-side FedAvg, and score-driven governance decisions.The model’s scores drive data-driven promotion and demotion decisions.Source: 2) Provider · p. 2

2) Provider:

  • A provider is the governed unit: its daily features are computed at one host, while party activity aggregates into provider features.The system uses provider-day aggregates and does not observe party-level data or another host’s data.Source: 2) Provider · p. 2, §III · p. 2

A. Federated Temporal Pairwise Ranking

DeFiRank simulates structurally heterogeneous institutions and trains a temporal pairwise ranker locally, aggregating model updates without sharing raw data, labels, or gradients.

A. Federated Temporal Pairwise Ranking

  • Five non-IID clients receive distinct validator-weight perspectives, so the same provider pair can receive opposite labels across institutions.The perspectives prioritize reciprocity, engagement, churn, volume, or interconnectivity; the partition is an experimental simulation of institutional heterogeneity.Source: A. Federated Temporal Pairwise Ranking · p. 3

B. Local Training

  • Training uses sliding windows of 13 governance features and generates pair labels by comparing providers’ governance scores within each day.The process yields 9,657,042 training pairs across all clients; yAB is 1 when provider A has higher governance quality than B, and 0 otherwise.Source: B. Local Training · p. 3
  • Local TemporalRankNet combines a two-layer LSTM over each provider’s 7-day sequence with RankNet pairwise ranking of same-day providers.The final LSTM state yields a 64-dimensional behavior embedding; pairwise binary cross-entropy trains the ranking model.Source: B. Local Training · p. 3
  • Each round broadcasts global weights, trains clients locally, and aggregates their updated weights via dataset-size-weighted FedAvg without sharing raw data, labels, or gradients.Training runs for 100 communication rounds, with the global model converging within the first 10.Source: B. Local Training · p. 3

C. Governance Decision Making

The governance pipeline converts daily provider scores into promotion and demotion decisions using a 30-day baseline, adaptive thresholds, and reliability checks.

C. Governance Decision Making

  • Daily scores feed promotion and demotion decisions after a 30-day baseline period.The pipeline first establishes baseline anchors, then evaluates providers daily and applies a reliability check to candidates.Source: C. Governance Decision Making · p. 3
  • Promotion thresholds and provider-count cutoffs adapt daily to network health and size, while low-activity featured providers can be dropped from consideration.The promotion threshold is clipped to [0.10, 0.60], and the cutoff to [3, 150], capped by featured-provider count.Source: A. Dataset · p. 4
  • Promotion candidates require qualifying runs of at least two consecutive days and an average ratio meeting the minimum reliability threshold.The baseline sets threshold anchors and the minimum outrank ratio used in this reliability check.Source: A. Dataset · p. 4

A. Dataset

  • The Canton Network dataset contains 965,894 provider-day samples with 28 behavioral features derived from transaction histories.The data covers 184 days; features capture transaction activity, engagement, churn, and network connectivity.Source: A. Dataset · pp. 3–4

1) Data Preprocessing:

The preprocessing and feature pipeline cleans provider records, selects governance and growth signals, and normalizes the resulting features for consistent scaling.

1) Data Preprocessing:

  • Missing provider or date identifiers are discarded, duplicate provider-date records retain the latest entry, numeric missing values become zero, and governance booleans are encoded as 0 or 1.Time-related columns are converted to standard datetime format.Source: 1) Data Preprocessing · p. 4

2) Feature Selection:

  • Thirteen features represent provider stability and governance quality, combining transaction and network behaviour with user-growth and engagement-depth measures.The selected base features are supplemented by volume, engagement, and user-growth trends computed from values seven days apart; trends are clipped to [−5, +5] and set to zero with fewer than seven days of history.Source: 2) Feature Selection · p. 4

3) Feature Normalization:

  • Features are z-score normalized, with the data’s mean and standard deviation saved to support consistent scaling across settings.The statistics are saved to feature mu.csv and feature sigma.csv.Source: 3) Feature Normalization · p. 4

4) Governance Score Computation:

A weighted governance score generates the pairwise training labels, while varied validator and client priorities create alternative perspectives on feature importance.

4) Governance Score Computation:

  • Because the dataset has no ranking column, a weighted linear combination of normalized features supplies governance scores for generating pairwise labels.The governance score is not used as a direct model input.Source: 4) Governance Score Computation · pp. 4–5
  • Feature weights begin from a base vector, receive small Gaussian perturbations, are floored at 10^-3, and are renormalized to sum to 1.Reciprocity rate has the highest base weight at 0.18; the perturbations simulate differing validator priorities.Source: 4) Governance Score Computation · p. 5
  • Federated clients use noisy, renormalized versions of five structural profiles emphasizing reciprocity, engagement, churn, volume, or interconnectivity.The emphasized dimension receives a base weight of 0.35–0.40, and the same provider pair can receive opposite labels from different clients.Source: 4) Governance Score Computation · p. 5

5) Pairwise Dataset Construction:

The method samples provider pairs within governance rounds to build a tractable pairwise dataset, then evaluates centralized and federated models with a separate date-level holdout.

5) Pairwise Dataset Construction:

  • Sampling per governance round is capped at min(5N_t, 5000), with featured-status restrictions removed so all three featured/non-featured pair types can occur.Exhaustive pairing is computationally infeasible for 965,894 samples.Source: 5) Pairwise Dataset Construction · p. 5
  • Pairs with governance-score differences below 10^-6 are discarded; remaining labels encode which provider has the higher score and use each provider’s full seven-day sequence.Each side has 91 feature columns, for 182 total plus the label column.Source: 5) Pairwise Dataset Construction · p. 5
  • The sampling procedure produces 9,657,042 record pairs across governance rounds.Source: 5) Pairwise Dataset Construction · p. 5

B. Experimental Setup

  • Ranking evaluation uses a separate date-level 80:20 holdout to prevent temporal leakage; centralized training runs 100 epochs, while federated training simulates five clients over 100 rounds.The federated clients have non-IID heterogeneity (Dirichlet α = 0.5) and client-specific structural weight profiles.Source: B. Experimental Setup · p. 5

C. Model Architecture

The models score providers from seven-day feature sequences and learn pairwise orderings; ranking quality is assessed both near the top of the list and across the full provider ordering.

C. Model Architecture

  • TemporalRankNet uses a two-layer LSTM, while AttentionRankNet uses four-head self-attention; both encode each provider’s seven-day, 13-feature sequence into a score for same-day pairwise comparison.Both optimize pairwise ranking loss over 9,657,042 training pairs.Source: C. Model Architecture · p. 5, Figure 3 · p. 6

D. Performance Metrics

1) Normalized Discounted Cumulative Gain:

2) Kendall

  • Kendall τ measures rank correlation across provider pairs, complementing NDCG’s focus on top-ranked precision with overall pairwise consistency.Its range is −1 to 1.Source: 2) Kendall · pp. 5–6

E. Baselines

The experiments compare centralized training on the full dataset with federated local training that shares model updates rather than raw transaction data. Crossing these regimes with LSTM-RankNet and Attention-RankNet gives four settings.

E. Baselines

  • The centralized regime trains on the full dataset at one server, while federated clients train locally and share only model updates.The federated setting does not expose raw transaction data.Source: E. Baselines · p. 6
  • The comparison tests both LSTM-RankNet and Attention-RankNet under centralized and federated training.This produces four settings: C-L, F-L, C-A, and F-A.Source: E. Baselines · p. 6
  • Centralized training has direct access to all provider data, unlike the federated regime.Source: E. Baselines · p. 6

F. Results

LSTM-RankNet’s federated results approach centralized ranking performance and match its Kendall τ, while Attention-RankNet’s NDCG is lower and less stable. Governance outcomes also differ with model-specific promotion anchors, while the demotion curves coincide.

F. Results

  • Federated LSTM-RankNet’s round-1 drop reflects conflicting client labels for provider pairs, and NDCG@k converges by round 6.The first aggregation reconciles local models before a stable global ranking emerges.Source: F. Results · p. 6
  • Federated LSTM-RankNet ends at NDCG@k (0.86, 0.87, 0.88) versus centralized (0.87, 0.91, 0.92), while both reach τ=0.70.Source: F. Results · p. 6
  • Centralized Attention-RankNet ends at NDCG values (0.75, 0.82, 0.84), and federated ends at (0.72, 0.74, 0.77) with τ=0.70.The centralized attention model’s NDCG fluctuates more than LSTM-RankNet’s; the federated attention model oscillates without a clear trend.Source: F. Results · p. 6
  • Attention-RankNet converges roughly five times slower than LSTM-RankNet under federation; its all-days attention may dilute 7-day trends, and its extra parameters likely contribute to instability.Source: F. Results · p. 6
  • Centralized demotion candidates split roughly two-thirds consistently poor and one-third borderline, while qualifying centralized promotion candidates had consistent signals over 2+ days.Source: F. Results · p. 6
  • Federated promotion uses stricter anchors, so fewer providers qualify than centrally, but qualifying providers have a higher average ratio.The centralized setting has lower promotion and minimum ratio anchors; each regime derives anchors from its own baseline period.Source: F. Results · p. 6
  • Centralized promotion ratios concentrate around 0.3–1.0, peaking near 0.65–0.7; federated ratios peak near 0.75–0.80 and have fewer candidates.The federated distribution is narrower and right-shifted, which the passage links to its stricter minimum ratio anchor.Source: F. Results · p. 7
  • Both models’ demotion curves are identical at the maximum Kbase = 150, while promotion thresholds differ by anchor and stay at τt = 0.10 after day 40.The passage attributes the post-day-40 minimum to declining network health and rapid network growth; both models show small peaks around days 87 and 131.Source: F. Results · pp. 7–8

V. CONCLUSION

DeFiRank replaces static manual epoch scoring with federated pairwise ranking, preserving institutional privacy while producing comparable ranking stability despite lower NDCG than centralized training.

V. CONCLUSION

  • NDCG@15: ≈0.87 vs. ≈0.92 for the centralized baseline, while Kendall τ: ≈0.70 for both regimes.The evaluation used a real-world blockchain dataset.Source: §V · p. 8
  • Across five non-IID clients, TemporalRankNet combines LSTM temporal encoding and RankNet pairwise ranking over 13 governance features and 9.6 million pairwise comparisons.Source: §V · p. 8
  • Federated governance produced 648 unique promotion candidates and 39 unique demotion candidates, with 100% promotion and 72% demotion consistency over 2+ consecutive days.A dynamic pipeline adaptively adjusts promotion and demotion thresholds in response to network health and size.Source: §V · p. 8
  • AttentionRankNet achieved comparable Kendall τ but consistently lower and less stable NDCG, motivating LSTM-RankNet for governance decisions.Source: §V · p. 8
  • Future work will explore tokenomic incentives, personalized federation for non-IID client heterogeneity, and formal differential privacy guarantees.Source: §V · p. 8
Loading 2610.12916v1…