Knowledge Graph-Driven Recommendations — Katariina Kari, Inter IKEA Systems

TL;DR

IKEA.com requires large-scale accessory recommendations (at least 10 per furniture item), making manual pairing impossible across thousands of variants.

Briefing Cornell Notes

Briefing

IKEA’s franchise network faces a scaling problem that manual merchandising can’t solve: in-store designers can curate a handful of accessory pairings per sofa, but IKEA.com must generate at least 10 accessory recommendations for nearly 2,000 sofa variants—creating tens of thousands of combinations across the catalog. The core fix being pursued is a knowledge-graph approach that encodes interior-design “why this goes with that” relationships as reusable business rules, so recommendations can be produced reliably and consistently at web scale.

In physical stores, the work is bounded. A showroom might display 20–25 sofas with a few accessory types—cushions, throws, and even practical add-ons like a lint roller—so interior designers manage roughly 75 combinations. Online, the same logic explodes: every sofa variant needs many accessory suggestions, and availability constraints (out of stock, region differences) can increase the effective complexity. Relying on “wisdom of the crowd” (cookie- and click-driven co-purchase statistics) is also increasingly brittle as privacy changes reduce usable data, and it often delivers results without the underlying rationale—sometimes producing dangerous or simply wrong pairings.

IKEA’s alternative centers on capturing expert knowledge rather than inferring it from anonymous behavior. The company’s hypothesis is that if recommendations include the “why” behind pairings—derived from interior designers’ mental models—then the system can deliver higher-quality, pre-vetted suggestions and potentially inspire customers with explanations. Those same structured relationships are also positioned as a foundation for other AI tasks, including computer vision: if a cushion is known to go with a sofa, image recognition systems can learn to identify those connected items in photos.

To operationalize this, IKEA is building an IKEA Knowledge Graph with a layered model for business-rule authoring and execution. At the ontology layer (the “green” layer), the graph defines the expressiveness of the rule system: product attributes (material, shape, color), and the types of relationships such as “accessory goes well with furniture,” including sub-relationships like “makes it comfortable” or “helps caring for it.” At the category layer (the “blue” layer), interior designers define concrete business rules by composing attribute-based conditions—for example, a “cushion for indoors” rule driven by indoor suitability and cushion shape, and an “indoor sofa” rule driven by sofa shape and indoor suitability. At the data layer, millions of product instances are connected automatically: a specific lint roller is linked to the “helps caring for” relationship, and all matching cushions and throws are generated via rule-driven combinations. The result is a recommendation stream for IKEA.com that can produce hundreds of accessory suggestions per sofa variant without handcrafting each pairing.

The system also needs ranking and filtering beyond rule generation. IKEA.com-side logic ranks recommendations using internal style matching and price-level constraints, with future layers potentially incorporating availability, sales optimization, and personalization. While the infrastructure investment is front-loaded, the rules are designed for reusability: once the pipeline and graph services are in place, new products and attributes can be integrated without heavy migration. The remaining challenge is making rule authoring practical for non-technical domain experts, since traditional ontology languages like OWL and early tooling like Protégé are not considered sufficient for interior designers. The approach being pursued aims to turn expert merchandising knowledge into a maintainable, scalable engine for online accessory recommendations across the franchise network.

Cornell Notes

IKEA is using a knowledge graph to generate accessory recommendations online in a way that manual merchandising and “wisdom of the crowd” can’t scale. The key idea is to encode interior designers’ “why this goes with that” mental models as business rules, then let the graph automatically connect matching products at web scale. The model is organized into three layers: an ontology layer that defines rule vocabulary (attributes and relationship types), a category layer where designers compose concrete rules (e.g., “cushion for indoors”), and a data layer where millions of product instances are linked automatically to produce recommendation streams. This structure supports quality control, consistent explanations, and reuse for other AI uses such as computer vision. Ranking and filtering can then apply additional constraints like style and price, while availability and personalization can be added later.

Why is IKEA’s online recommendation problem harder than in-store merchandising?

In stores, designers curate a limited set of displays: roughly 20–25 sofa variants are shown, and each sofa gets only a few accessory types (like cushions, throws, and even a lint roller). That keeps the number of combinations manageable (about 75 combinations). Online, IKEA.com must handle nearly 2,000 sofa variants, and the UI requires at least 10 accessory recommendations per furniture item—often more when items are out of stock or unavailable in certain markets. Manually connecting each furniture to each accessory becomes infeasible, creating tens of thousands of combinations just for sofas, and far more across the full catalog.

What’s wrong with relying on “wisdom of the crowd” for recommendations?

Crowd-based recommendations depend on large volumes of behavioral data (clicks, purchases, cookies). As privacy controls increase and more users opt out, the data driving statistical co-occurrence declines. Even when data exists, the approach tends to provide the recommendation without the underlying rationale (“why”), which limits customer trust and makes mistakes harder to vet. A cited example involved a person seeking a nebulizer and receiving an inappropriate recommendation from a well-known platform—illustrating that incorrect pairings can be difficult to detect after the fact.

How does the knowledge graph capture the “why” behind accessory pairings?

The system encodes interior-design relationships as business rules. The ontology layer defines product attributes (material, shape, color) and relationship types such as “accessory goes well with furniture,” including more specific properties like “makes it comfortable” or “helps caring for it.” The category layer then composes concrete rules from those attributes (e.g., “cushion for indoors” based on indoor suitability and cushion shape, and “indoor sofa” based on sofa shape and indoor suitability). Finally, the data layer applies these rules to actual product instances so that specific items (like a particular lint roller) are linked to the appropriate relationship and all matching accessories are generated automatically.

How do the three knowledge-graph layers work together to produce recommendations?

Ontology (green): defines the rule vocabulary—attributes and relationship semantics. Category (blue): stores the actual business rules created by designers by combining attribute conditions (e.g., indoor suitability + shape). Data (bottom): instantiates the rules across millions of products. For a given sofa, the system combines the left-side accessory rule outputs (matching cushions/throws/lint rollers) with the right-side furniture rule outputs, effectively performing a Cartesian-style combination to generate hundreds of accessory recommendations per furniture variant. This is repeated across furniture types to feed IKEA.com.

How are recommendations ranked and filtered after rule-based generation?

Rule-based generation creates candidate accessory sets; ranking and filtering apply additional constraints. IKEA uses an internal style system to match styles between furniture and accessories, and it also matches price level to keep recommendations aligned with customer budget. Later layers can incorporate availability (stock checks), sales or margin optimization, and potentially personalization—some of which may use knowledge-graph signals and/or machine learning.

What makes the approach practical for domain experts like interior designers?

The graph is designed for reusability and maintainability: adding new information (like a new material or style) is described as not requiring heavy migration, because the data model is dynamic and new attributes can flow through established pipelines. The remaining usability challenge is authoring: ontology languages such as OWL and tools like Protégé are not considered suitable for non-technical users. IKEA is therefore working toward tooling that lets interior designers define and refine rules through a more accessible workflow.

Review Questions

How does encoding “why” as business rules improve recommendation quality compared with co-purchase statistics?
Describe what changes when moving from the ontology layer to the category layer to the data layer in IKEA’s rule model.
What additional constraints beyond compatibility are needed to make recommendations useful on IKEA.com (e.g., style, price, availability)?

Key Points

1
IKEA.com requires large-scale accessory recommendations (at least 10 per furniture item), making manual pairing impossible across thousands of variants.
2
Cookie- and click-driven “wisdom of the crowd” weakens as privacy reduces data and often lacks an explicit rationale for recommendations.
3
IKEA’s approach encodes interior designers’ “why this goes with that” as business rules in a knowledge graph to produce pre-vetted, higher-quality pairings.
4
The knowledge graph is structured into ontology (rule vocabulary), category (designer-authored rules), and data (automatic product instantiation) layers.
5
Rule-based generation can create hundreds of accessory candidates per sofa variant by combining attribute-driven accessory rules with furniture rules.
6
Recommendation ranking uses additional signals such as internal style matching and price-level constraints, with future layers for availability, sales optimization, and personalization.
7
A major implementation goal is making rule authoring usable for non-technical interior designers, since standard ontology tooling is not considered sufficient.

Highlights

The physical store can manage accessory pairings because only a small subset of variants is displayed; online must cover nearly 2,000 sofa variants and generate 10+ accessory suggestions per item.

The system targets the missing piece in crowd-based recommendations: an explicit “why,” derived from expert interior-design mental models.

Business rules are composed from product attributes (like indoor suitability and shape) and then applied automatically to millions of product instances.

Ranking is separated from compatibility: style and price constraints refine the candidate set produced by the knowledge graph.

The infrastructure is built once for reuse, while new materials or styles can be added without heavy migration—assuming the data pipeline supports it.

Topics

Knowledge Graph Recommendations
Business Rules
Ontology Layer
Interior Design
E-commerce Ranking

Mentioned

Katariina Kari
Adam
OWL
KGC

Knowledge Graph-Driven Recommendations — Katariina Kari, Inter IKEA Systems | KGC 2023