Building explainable and trustworthy recommendation systems: What we learned from IKEA at KGC 2023

Use Cases

Pauline Leoncio



Reading time: 9 - 11 minutes

building explainable trustworthy recommendation systems IKEA


In this blog post, we dive into how knowledge graphs play an important role in IKEA's recommendation systems, based on our experience attending two presentations by IKEA at the 2023 Knowledge Graph Conference.


Building explainable and trustworthy recommendation systems: What we learned from IKEA at KGC 2023


Can knowledge graphs enhance existing recommendation systems and make them more accurate, explainable and trustworthy?


This question, along with other topics around knowledge graphs and generative AI were central to the discussion at the annual Knowledge Graph Conference 2023, held earlier this year in New York, USA. KGC 2023 is one of the largest international conferences around knowledge graphs and the semantic web that brings together researchers and industry leaders from the knowledge graph community, online and in-person, to discuss its changes, trends and new applications. 


In this blog post, we’ll recap what we learned from both presentations and explore the knowledge graph’s place in an increasingly AI world. Keep reading!


Table of contents 



What is a recommendation system? 

This year, we had the pleasure of attending as both a sponsor and presenter with the opportunity to listen in on many great presentations by our peers and partners in the industry. Some of the presentations we attended included the insightful sessions held by Inter IKEA Systems B.V.’s lead ontologist, Katariina Kari, and business owner of the IKEA knowledge graph, Adam Keresztes, who both delved into how IKEA created a knowledge graph to improve its recommendation systems.


Before diving into the presentations, here's a quick primer on the recommendation system and a review of its three main approaches. 


Recommendation or recommender systems are commonly realized using machine learning and encompass software tools and techniques that generate recommendations to users about products or content based on data such as user behavior and historical preferences.


Once you’re aware of what it is and how it works, you’ll find that recommendation systems are at play on almost every website or application you come across—from e-commerce shops and news and media platforms to streaming services like Netflix. This system functions to provide precise predictions that closely align with a user's preferences and intentions, which enhances their experience on the site or app by tailoring it to their individual preferences. 


Another advantage of this system is its ability to prompt users to engage in particular actions, such as making a purchase or clicking on a link. It’s particularly effective in driving action because it offers users options that are already aligned with their interests, amidst a sea of choices.


Types of recommender systems 

There are three main approaches to designing recommendation systems: Collaborative Filtering, Content-based Filtering and Hybrid Recommendation Systems. While we’ll provide an overview of these three main types, know that we’re only scratching the surface of what a recommendation system can do and how it can be applied. For example, whether we realize it or not, there are numerous applications that we come across in our daily lives, such as with music, hospitality and research. 


Collaborative filtering

The collaborative filtering method is the most widely used recommendation system. It relies on the historical behavior and interactions of a collective group of users with a site and/or products to make predictions about the future interests of an individual user. The core principle of this method is that, based on various factors, the system predicts a user's potential interest in items based on content that users similar to them—whether in proximity, user interaction or previous purchases—have shown interest in.


For example, if User A and User B both positively rate the movie Spiderman, and User B also positively rates the movie Iron Man, the assumption would be that User A would also like the movie Iron Man. Another thing to note is that similar opinions on a specific topic or category, in this case, superhero movies, would also be used to indicate similar interests in a different topic or category, such as crime movies. 


Content-based filtering

The content-based filtering method is based on a user’s profile and past interactions and the description of an item or content. This information is used to generate recommendations for that user, like an attribute-matching system. For example, if User A has watched three superhero movies in the past, then the system presumes that they would be interested in watching another superhero movie. 


Hybrid recommendation systems

The hybrid recommendation system combines two or more methods together, such as the collaborative filtering and content-based approach, to offer more accuracy in its recommendations. By leveraging the strengths of both collaborative and content-based methods, the hybrid system can overcome the problems that the individual approaches face when used alone. For instance, Netflix’s recommender system is comprised of multiple algorithms that assess factors such as trends, user profile, user behavior and highly-rated content to construct an ultra-personalized homepage that reflects precisely the kinds of TV shows and films each viewer would like to watch. 


Limitations of “The Wisdom of the Crowd” 

In Kari’s presentation at KGC 2023 titled, Knowledge Graph-Powered Recommendations, she highlighted the collaborative filtering method, referring to it as the ‘wisdom of the crowd’ approach, and suggests that while the wisdom of the crowd is today’s industry standard for recommendation systems, it comes with many limitations that prevent it from ever being a truly reliable approach.


Data quality

One of the main challenges Kari shared regarding the 'wisdom of the crowd' approach is that there’s overall less data to analyze and power these recommendations. The General Data Protection Regulation (GDPR) in Europe, set into effect in 2018, grants web visitors the option to opt-in to third-party website cookies (which are data stored on a user’s device to track their online behavior), rather than the automatic acceptance that was once the standard. As a result, fewer people accept these third-party cookies, resulting in less users to glean information from.


On the other hand, Kari explained that when customers do accept third-party cookies, they perceive the recommendations they receive as creepy, evoking concerns over being digitally stalked by a company. Too close of an analysis of user data across websites can also lead to suspicion and discomfort, and ultimately elicit distrust in the company and its recommendations. 


Can’t explain the “why”

According to Kari, another limitation is that the wisdom of the crowd can’t provide the why to the recommendations it gives. It can suggest what item you should buy for your home but offers no explicit explanation for why this product would be beneficial for you. Being unable to explicitly describe the reasoning behind each recommendation may lead customers to question the quality of its suggestions. 


Can generate mistakes (that can be deadly)

Kari shared an unsettling story about the time her colleague fell sick with the flu and sought out a nebulizer on a very popular global e-commerce platform. Her colleague was shocked to discover that one of the items being recommended as an accompaniment to the nebulizer was hydrogen peroxide, which paired together is obviously a lethal combo. 


While there’s much information to derive from the data captured from online user behavior, there’s still an opportunity to create more value for users through alternative approaches to recommendations, one that users can better trust and verify.


Manual tagging

Another example of a method used to drive recommendations at IKEA is manual tagging, which Keresztes shared in his presentation titled, Knowledge-Graph-enhanced Product Catalogues. When developing product catalogs at IKEA, it's necessary to capture information like size, color and pricing, as well as qualities such as suitability for families and different sizes of homes. While the former information is easier to come by, the latter is only known by subject matter experts (i.e., IKEA's interior designers) who have intimate knowledge of the product, interior design and different home lifestyles. 


Keresztes said that they had no simple mechanism for capturing this information initially, so they began manual tagging, which means individually labeling a product with attributes in a system to ensure each product is tagged with attributes useful for customers to know, including the attributes that are only known by the subject matter experts. This process involved the laborious task of tagging products individually without being able to link them to an overarching model.


An example he gave was regarding the IKEA TINGBY side table. To the everyday customer, the TINGBY is a minimal-looking side table with two shelves, casters and rounded corners. Only a subject matter expert would be able to identify that the TINGBY side table is ideal for small spaces and families because its casters make it mobile, which is good when needing to make room in a small space, the shelf is a space-saver because it can be used to hold more items than a typical side table, also making it ideal for small spaces, and the rounded corners of the TINGBY make it a safer option for homes with children. With manual tagging, the subject matter expert holding this knowledge would then tag the TINGBY as: suitable for small homes and suitable for families. Customers looking for products that fall into these categories would then be presented with the TINGBY and other related products.


 [Knowledge-Graph-enhanced Product Catalogues, Adam Keresztes, KGC 2023]


However, there are still many difficulties with developing catalogs using this method.



First, there’s the issue of quality, said Keresztes. Differences in human experience and perspective can largely shape one's interpretation and contextual understanding of a product, leading to potential inconsistencies between different subject matter experts. Quality can also fluctuate between seniority level and expertise. 



When subject matter experts go on leave for illness, parental leave or permanently exit the company, they take with them a wealth of knowledge that only they may have, therefore making it difficult to sustain the same quality of knowledge and philosophy over the years. 



IKEA has over 70,000 products in its inventory, with new products being developed every year and multiple times a year. The number of products that exist makes manually tagging each product a cumbersome task—and it’s especially difficult when new circumstances arise, such as a global pandemic, for example, because it requires updating the attributes of existing products to account for changes in home and lifestyle needs. According to Keresztes, it’s too gargantuan of a task, and makes no sense, to backfill and re-tag older products, even though it means many products wouldn't be as up-to-date as they should be.



Keresztes' last point is around transparency, which is also the central question for all these recommendation systems. Can this system or method explain the why along with the what? Without proper data, it’s impossible to explain why a side table like the TINGBY  is better suited for customers with smaller homes, which is crucial when trying to build trust and persuade customers to take action. 


Seeing the limitations of both the ‘wisdom of the crowd’ and manual tagging, the team at Inter IKEA Systems B.V. wanted a new way to capture and share the specialized knowledge held only by subject matter experts to better provide customers with a personalized experience they can rely on. 

How IKEA makes trustworthy recommendations

Rather than relying on ML solutions based on statistical methods (that can lead to false positives or inaccurate recommendations), or the manual tagging of attributes (laborious and unsustainable), Kari reveals that IKEA built its own knowledge graph, powered by metaphactory, to drive smarter recommendations. A semantic knowledge graph explicitly defines the relations between entities—in IKEA’s case, products—and the contextual meaning behind these relations through a semantic model, also known as an ontology or semantic data model. This semantic model is shaped by subject matter experts who define the relations, attributes and business-specific terminology of each entity. 


Inside physical IKEA stores, interior designers can pair complementary products together and place add-ons nearby so customers can see how their desired product and IKEA’s recommended items work together. Capturing interior designer mental models as business rules in a knowledge graph enables IKEA to harness the wisdom of its subject matter experts with technology using semantics, and replicate this experience online. 


Interior designers are aware of the style, functionality and real-world applications that need to be considered when purchasing furniture or home decor, which aren’t obvious from the outset or to the untrained eye. This knowledge can prevent potential mistakes, such as labeling a side table with sharp corners 'suitable for families’ when it is actually an unsafe option for homes with babies and young children.  


IKEA’s knowledge graph also follows the mental models and concepts that customers naturally think in, and take into consideration the different life aspects important to customers when shopping for furniture, rather than the rigid and narrow product categorizations that disregard these essential factors. Often a customer does not just look for a sofa of a certain size and color, they may search for a sofa that fits a small space, is robust and has washable covers.


Business rules are often captured in people’s minds (and cannot be documented) or are captured in applications (that cannot be extended for reuse across the company). Being able to structure data in this way enables IKEA’s recommender system to make recommendations that are thoughtful, customer-focused and true-to-life, as well as be extendable to other applications such as image recognition.

IKEA’s Knowledge Graph

In her presentation, Kari demonstrated how IKEA sets up business rules for indoor sofa accessories. In the visual below, you’ll see a pyramid representing the three layers of the IKEA knowledge graph and how it’s used in practice. 


The top layer

[Knowledge Graph-Powered Recommendations, Katariina Kari, KGC 2023]


The top or green layer of IKEA’s knowledge graph can be understood as the ‘ontology layer’ in a semantic knowledge graph. According to Kari, this layer is where different kinds of product attributes are captured, such as product material, shape and color. It also captures the definition of business rules for accessories and furniture. For example, defining these attributes and properties, such as explicitly indicating that accessories go well with furniture because they either make it more comfortable or are simply well-suited for it, helps to formulate a business rule.

The middle layer

[Knowledge Graph-Powered Recommendations, Katariina Kari, KGC 2023]


The middle or blue layer is what can also be referred to as the vocabulary or taxonomy layer, and it’s where actual business rules are defined by IKEA’s interior designers. An interior designer may define that a specific cushion shape, such as a square-shaped cushion, should link with the product attribute of being suitable for indoors. They may also define that an L-shaped sofa is also suitable for inside the home and connect these separate rules. The resulting rules would be that cushions suitable for indoors also pair well with indoor sofas, and indoor cushions make indoor sofas comfortable. Since not all IKEA cushions are made for the indoors, this type of product knowledge is something that customers aren't inherently privy to. These are examples of the many rules that can be created between just two types of products. 

The bottom layer

[Knowledge Graph-Powered Recommendations, Katariina Kari, KGC 2023]


The final, bottom layer of the pyramid is the instance data level, where millions of instances (i.e., IKEA products) exist, and where the connections between these instances are explicitly defined. Kari explained that on this level, general rules only need to be expressed once through product attributes. Due to the existing business rules defined in the previous, interior designers wouldn't need to manually tag or pair accessories and products together—these rules would be automatically applied across all instances. For example, as expressed by the rules above, it would be defined that the SVARTHÖ cushion makes the SÖDERHAMN sofa more comfortable.


[Knowledge Graph-Powered Recommendations, Katariina Kari, KGC 2023]


Kari and Keresztes' presentations reveal how knowledge graphs not only complement but also enhance recommendations by improving their accuracy and trustworthiness, which is achieved by capturing the context revolving around each individual item only known by subject matter experts. By explicitly expressing knowledge through a knowledge graph, you can capture the intricacies and nuances of real-world objects and concepts. This specialized knowledge can then be used to power various tools and applications, such as enhancing the reliability of and adding explainability to data-driven AI and ML solutions. In IKEA's case, customers will be able to trust that what's being recommended to them truly makes sense for their lifestyle and home. 


Kari concluded her presentation by suggesting that more tool definitions for non-technical users are needed to encourage collaboration from subject matter experts of varying technical expertise and that there should be widespread adoption of KGs across different industries. 


We thank both Kari and Keresztes for their engaging presentations, along with the other presenters and organizers of KGC 2023. We look forward to convening once again at next year's event!


Try it yourself 

Learn how your enterprise can enhance and complement your recommendation system with a semantic knowledge graph by seeing how our knowledge graph platform, metaphactory, works in practice. You can get started with metaphactory today using our free trial.  


metaphactory is an industry-leading enterprise knowledge graph platform that helps you transform data into consumable, contextual and actionable knowledge. Our low-code, FAIR Data platform simplifies capturing and organizing domain expertise, extracting actionable insights from your data and sharing knowledge across the enterprise. 

metaphactory includes innovative features and tools for: 


  • Knowledge management — manage your knowledge graph assets - incl. ontologies, vocabularies, data catalogs and instance data -  in one place
  • Low-code application building — build easy-to-configure applications that fit your enterprise requirements 
  • User-friendly interface — users of any level of technical experience can interact with your data through a user-friendly interface that includes search, visualization, discovery & exploration, authoring and end-user knowledge management 


Get your free trial of metaphactory


Pauline Leoncio

Pauline Leoncio is an experienced copywriter and content marketer with over 6+ years in marketing. She's developed content plans and creative marketing material for growing B2B and B2C tech companies and covers a range of topics including finance, advanced tech, semantic web, food, art & culture and more.