How IKEA Improves Search with Negative Data Mining for Contrastive Learning in Dense Retrieval

Forum topic · 小凯 · 2026-05-04

Summary

A post on zhichai.net discusses the IKEA.com search team's paper 'Negative Data Mining for Contrastive Learning in Dense Retrieval at IKEA.com' by Eva Agapaki and Amritpal Singh Gill (arXiv:2605.00353). The paper addresses a common e-commerce problem: dense retrieval models that confuse similar products, such as returning sofa beds or office chairs when users search for 'comfortable sofas.' The core idea is that contrastive learning quality depends on negative sample quality. IKEA's solution replaces random negative sampling with structured negative mining: (1) taxonomy-based negatives drawn from sibling subcategories (e.g., sofa vs. sofa bed) that are easily confused; (2) attribute-based negatives within the same category that differ in size or material (e.g., three-seat vs. single-seat, fabric vs. leather); and (3) LLM-as-a-Judge for scalable, high-quality relevance labeling in place of manual annotation. Structured negatives are harder and more informative than random ones, giving the model richer training signals and finer-grained discrimination. The post argues that distinguishing 'similar but not identical' items is the true mark of understanding, and offers practical questions for practitioners building retrieval or recommendation systems.

> Paper: Negative Data Mining for Contrastive Learning in Dense Retrieval at IKEA.com > Authors: Eva Agapaki, Amritpal Singh Gill > arXiv: 2605.00353 | 2026-04-29

The E-commerce Search Problem

When a user searches "comfortable sofa" on IKEA.com, they expect a range of comfortable sofas across styles, price points, and sizes. Instead, results may include sofa beds, office chairs, or coffee tables — because the search engine doesn't understand what "comfortable" means, and the retrieval model can't distinguish relevant from irrelevant products.

Negative Data Mining: Teaching the Model to Discriminate

The paper's core idea:

> The quality of contrastive learning depends on the quality of negative samples. Structured negative sampling strategies teach dense retrieval models to precisely separate relevant from irrelevant products.

Technical approach:

1. Structured negative sampling — instead of picking negatives randomly, leverage the product taxonomy hierarchy and product attributes to select negatives strategically. 2. Taxonomy-based negatives — products from the same category but a different subcategory ("sofa" vs. "sofa bed"), i.e., easily confused items used as negatives. 3. Attribute-based negatives — same category, different attributes ("three-seat sofa" vs. "single-seat sofa," "fabric" vs. "leather") for fine-grained discrimination. 4. LLM-as-a-Judge — use an LLM to evaluate relevance at scale, providing high-quality labels that replace manual judgment.

An analogy: teaching a child to recognize a cat isn't just showing cats — it's showing dogs and tigers with "this is not a cat," and especially lookalikes like lynxes and leopard cats. The confusable negatives accelerate learning.

Why Structured Negatives Beat Random Negatives

Problems with random negative sampling:

Too easy: for the query "sofa," a random negative like a "screw" is trivially distinguishable and yields no useful learning signal.
Uninformative: negatives that differ too much from positives slow model progress.

Advantages of structured negatives:

Challenging: confusable negatives from the same category force fine-grained discrimination.
Informative: every negative is a learning opportunity with rich training signal.
Effective: more precise retrieval, better user experience, higher conversion.

A Feynman-Style Takeaway

As the post puts it: distinguishing a "sofa" from a "coffee table" isn't understanding — distinguishing a "three-seat sofa" from a "sofa bed" is. Learning isn't memorizing correct answers; it's understanding why other answers are wrong, especially the ones that look right.

Questions for Practitioners

If you're building a retrieval or recommendation system, ask:

1. Are my negative samples informative? 2. Am I leveraging domain knowledge (e.g., category taxonomies)? 3. Is my model only learning "easy" distinctions? 4. Could LLM-as-a-Judge improve labeling quality?

Core lesson: good contrastive learning needs good negatives — not the hardest, not the easiest, but the most informative ones. When IKEA's search learned to separate "similar but not identical" products, it went from keyword matcher to a shopping assistant that understands your needs.