EARS-XTSK: Privacy-Preserving Global Explainability in Cross-Silo Federated Two-Tower Recommendation Systems via Server-Side TSK Fuzzy Rule Distribution

Recommender systems form the both the commercial and the computational backbone of personalization in modern digital ecosystems—from video streaming and e-commerce to news, finance, and social platforms. Yet, the data centralization they require conflicts with privacy regulations such as GDPR and KVKK, which are emphasized more than ever by both the policymakers and the end users, motivating the adoption of privacy-first and transparent solutions. In the privacy context, a prominent approach is Federated Learning (FL). Yet, on its own, it lacks the transparency and interpretability that a centralized solution offers under the same circumstances. We present EARS-XTSK, a privacy-preserving global explainability framework for cross-silo federated recommendation. Our industry setting comprises two real-life data providers in Türkiye: (i) a large online forum representing the user metadata side, and (ii) a commercial streaming platform providing item/content and interaction data. By design, raw user data never leaves partner premises. To enable the global FED-XAI layer without direct data access, we construct a faithful synthetic dataset conforming to the partners’ schema, feature definitions, and scale, and we build an end-to-end clone of their pipelines to de-risk integration. Local sites train Two-Tower retrieval models (user tower on demographics/geo, item tower on multilingual text and metadata) and participate in NVFLARE-based FedAvg aggregation. As our core contribution, we introduce a server-side explainability engine that fuses Grad×Input saliency with a Takagi–Sugeno–Kang (TSK) fuzzy-rule layer to produce interpretable global explanations over item-side labels only (privacy-strict). We provide a complete, reproducible pipeline: dummy data generation (OMDb, SBERT, fastText), Two-Tower embedding/training, federated orchestration (NVFLARE), and Fed-XAI evaluation with visualization. Prototype runs demonstrate faithful global explanations without exposing user identifiers or raw features. In deployment, the same FED-XAI is practically guaranteed to work with the partners’ real local Two-Tower models, since all upstream interfaces are schema-compatible.