Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Optimizing Product Deduplication in E-Commerce with Multimodal Embeddings

Aysenur Kulunk, Berk Taskin, M. Furkan Eseoglu, H. Bahadir Sahin

TL;DR

The paper tackles duplicate product listings in large Turkish e-commerce by building a domain-specific, multimodal deduplication system. It combines a Turkish BERTurk-based text encoder with a Masked AutoEncoder–based image encoder to produce compact 128-dimensional embeddings, and uses a dedicated decider that fuses text and image vectors for fast, category-agnostic classification. Through Milvus vector search with IVF_FLAT indexing, the approach achieves a macro-F1 of 0.90, outperforming a strong third-party baseline, while maintaining low memory and latency suitable for hundreds of millions of items. The work demonstrates scalable, efficient, and accurate deduplication with potential deployment at-scale (116M product vectors daily) and outlines future directions for richer multimodal integration and language expansion.

Abstract

In large scale e-commerce marketplaces, duplicate product listings frequently cause consumer confusion and operational inefficiencies, degrading trust on the platform and increasing costs. Traditional keyword-based search methodologies falter in accurately identifying duplicates due to their reliance on exact textual matches, neglecting semantic similarities inherent in product titles. To address these challenges, we introduce a scalable, multimodal product deduplication designed specifically for the e-commerce domain. Our approach employs a domain-specific text model grounded in BERT architecture in conjunction with MaskedAutoEncoders for image representations. Both of these architectures are augmented with dimensionality reduction techniques to produce compact 128-dimensional embeddings without significant information loss. Complementing this, we also developed a novel decider model that leverages both text and image vectors. By integrating these feature extraction mechanisms with Milvus, an optimized vector database, our system can facilitate efficient and high-precision similarity searches across extensive product catalogs exceeding 200 million items with just 100GB of system RAM consumption. Empirical evaluations demonstrate that our matching system achieves a macro-average F1 score of 0.90, outperforming third-party solutions which attain an F1 score of 0.83. Our findings show the potential of combining domain-specific adaptations with state-of-the-art machine learning techniques to mitigate duplicate listings in large-scale e-commerce environments.

Optimizing Product Deduplication in E-Commerce with Multimodal Embeddings

TL;DR

The paper tackles duplicate product listings in large Turkish e-commerce by building a domain-specific, multimodal deduplication system. It combines a Turkish BERTurk-based text encoder with a Masked AutoEncoder–based image encoder to produce compact 128-dimensional embeddings, and uses a dedicated decider that fuses text and image vectors for fast, category-agnostic classification. Through Milvus vector search with IVF_FLAT indexing, the approach achieves a macro-F1 of 0.90, outperforming a strong third-party baseline, while maintaining low memory and latency suitable for hundreds of millions of items. The work demonstrates scalable, efficient, and accurate deduplication with potential deployment at-scale (116M product vectors daily) and outlines future directions for richer multimodal integration and language expansion.

Abstract

In large scale e-commerce marketplaces, duplicate product listings frequently cause consumer confusion and operational inefficiencies, degrading trust on the platform and increasing costs. Traditional keyword-based search methodologies falter in accurately identifying duplicates due to their reliance on exact textual matches, neglecting semantic similarities inherent in product titles. To address these challenges, we introduce a scalable, multimodal product deduplication designed specifically for the e-commerce domain. Our approach employs a domain-specific text model grounded in BERT architecture in conjunction with MaskedAutoEncoders for image representations. Both of these architectures are augmented with dimensionality reduction techniques to produce compact 128-dimensional embeddings without significant information loss. Complementing this, we also developed a novel decider model that leverages both text and image vectors. By integrating these feature extraction mechanisms with Milvus, an optimized vector database, our system can facilitate efficient and high-precision similarity searches across extensive product catalogs exceeding 200 million items with just 100GB of system RAM consumption. Empirical evaluations demonstrate that our matching system achieves a macro-average F1 score of 0.90, outperforming third-party solutions which attain an F1 score of 0.83. Our findings show the potential of combining domain-specific adaptations with state-of-the-art machine learning techniques to mitigate duplicate listings in large-scale e-commerce environments.

Paper Structure

This paper contains 26 sections, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Text and Image Representation
  4. Foundational Multimodal Models and Our Approach
  5. Product Deduplication Approaches
  6. Theoretical Analysis
  7. Methodology
  8. Feature Extraction and Dimensionality Reduction
  9. Text Feature Models
  10. Visual Feature Models
  11. Domain Adaptation
  12. Product Matching Pipeline
  13. Experiments & Results
  14. Experimental Setup
  15. Dataset
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Updated version of text model
  • Figure 2: (a) Second iteration of the image model P# denotes the selected patches, M# denotes the masked patches (b) Final iteration of the image model P# denotes the selected patches, C is the center patch, Xr is resized image
  • Figure 3: Final system design for product deduplication.
  • Figure 4: Scale differences of a product. (a) image does not have a white border, (b) image has a white border