Centralized item ranking in online marketplaces compromises user privacy and is vulnerable to manipulation. The introduction of a federated, reputation-based ranking system preserves privacy, ensures fairness, and delivers robust and effective rankings.

The growth of online marketplaces has transformed consumer experiences, offering diverse products aggregated from multiple sellers. However, the centralized nature of these platforms poses significant challenges, particularly in maintaining user privacy and defending against ranking manipulation. Addressing these issues, in a study conducted in cooperation with Guilherme Ramos and Mirko Marras, we present a novel framework for Robust Privacy-Preserving Federated Item Ranking, introducing a reputation-based mechanism to ensure fairness, effectiveness, and privacy preservation.

Traditional online marketplaces aggregate user ratings and preferences to present item rankings. While effective, this approach requires collecting user data from seller platforms, raising privacy concerns and compliance challenges (e.g., GDPR). Furthermore, centralized systems are vulnerable to manipulation through push or nuke attacks, where sellers artificially alter item rankings.

To address these, we proposes a federated learning paradigm where:

• User ratings remain private within seller platforms.
• Each seller generates its rankings independently.
• The marketplace aggregates rankings using a reputation-weighted system.

The proposed framework

The core of this framework lies in combining federated learning with reputation-based weighting. Key innovations include:

1. Non-personalized ranking aggregation. The framework focuses on non-personalized rankings, targeting scenarios where user preferences are unavailable (e.g., anonymous browsing).
2. Reputation-based weighting. Sellers’ contributions to the final ranking are weighted by their reputation. Reputation scores are computed based on the consistency of a seller’s rankings compared to others, penalizing discrepancies to deter manipulation.
3. Privacy preservation. Only aggregated rankings and metadata (e.g., item count) are shared with the marketplace, ensuring no direct access to user ratings or preference patterns.

Methodology

The study evaluates three aggregation strategies:

• Naïve Arithmetic Average (N-AA): Simple averaging of seller rankings.
• Weighted Arithmetic Average (W-AA): Weighting rankings by the number of user ratings per seller.
• Reputation-Based Aggregation (REP): Weighting rankings using computed seller reputations and user rating counts.

The computational complexity of REP is analyzed to ensure scalability in large-scale marketplaces.

Experimental validation

Using real-world datasets, the researchers assessed the framework’s effectiveness and robustness. Key findings include:

• Effectiveness. REP consistently outperformed N-AA and W-AA in producing rankings that closely matched a hypothetical ground truth (complete access to user ratings).
• Seller Satisfaction. REP achieved higher seller satisfaction by aligning marketplace rankings more closely with individual seller contributions.
• Robustness. REP demonstrated resilience against push and nuke attacks, significantly mitigating their impact through reputation-based penalties for outlier rankings.

Applications and future directions

This framework is valuable for marketplaces balancing privacy concerns with the need for effective recommendations. Future work will focus on:

• Incorporating real-world datasets with explicit seller-item associations.
• Extending the model to dynamic marketplaces with evolving seller reputations.
• Addressing scalability in scenarios with thousands of sellers and items.