A Multi-Modal Entity Resolution Architecture For Product Catalog Deduplication and Attribute Enrichment Across Heterogeneous E-Commerce Sources

Abstract

Reconciling product records that describe the same real-world item across merchant feeds, marketplace listings, and supplier catalogs remains one of the most persistent data quality problems in e-commerce systems. Heterogeneity in schema, terminology, image quality, and completeness makes naive key-based matching unreliable at scale, while the combinatorial cost of pairwise comparison across millions of listings makes exhaustive matching computationally infeasible. We propose a multi-modal entity resolution architecture that combines text-based attribute normalization, image-based visual similarity matching, and a fusion/scoring layer to resolve product-level entities across heterogeneous sources. The architecture further incorporates pricing-signal reconciliation as a corroborating matching feature and a confidence-based human-in-the-loop review mechanism for ambiguous cases that fall inside an uncertain decision band. We discuss blocking and candidate-generation strategies suited to catalog-scale deduplication, compare matching-signal types and their tradeoffs, and outline a taxonomy of blocking approaches applicable to production catalog systems. We frame this as an architectural synthesis grounded in production experience rather than a benchmarked empirical study, and discuss precision/recall tradeoffs, failure modes, and open research directions. We conclude that no single signal, textual, visual, or price-based, is sufficient in isolation, and that principled fusion combined with selective human review is the most durable path to reliable, explainable catalog deduplication at scale.