Skip and go to main content

Digital Marketing July 02, 2018

Personalise your digital product offering with a Recommender System

Recommender system header

Finding a product that fully meets the personal needs of every consumer is difficult – not to mention misplaced targeting that haunts customers for weeks after an online search. This is not exactly pleasant for consumers, but also not good for webshops, which miss out on a close connection and bond with both existing and potential customers. This can be fixed by providing personalised recommendations, both for products that are very similar to the initial search and for additional and complementary products. The solution? A Recommender System.


There is a solution that will help consumers easily find the product they are looking for. By leveraging big data and machine learning, we can recommend products that are fully tailored on a personal level and specifically adapted to the tastes and expectations of the customer. These product recommendations can be generated in two flavours both with a different purpose:

  1. Predicting alternative products, personalised and adapted to the customer’s wishes (one-on-one communication, algorithmically and no business rules). This includes offering an equivalent but different product based on your search.
    – This solution helps enhance the user experience. As a result, it is quicker and easier for the customer to find the product he or she is looking for.
  2. Predicting complementary products, also personalised and adapted to the customer’s wishes (one-on-one communication, algorithmically and no business rules). This includes recommending additional items that complement the initial product and are often sold, also known as Frequently Bought Togethers (FBTs).
    – This solution is referred to as cross-selling, i.e. trying to sell other products and is best applied during the checkout process or after a purchase. Due to the fact that these complementary product recommendations are generated on a personal level, more items will be sold hence the average order value will increase.


Building a customised Recommender System makes it possible to integrate the two solutions into the same system. The Recommender System uses machine learning and raw clickstream data to come up with recommendations in real time. So, what can you achieve with it? One-on-one communication with a personalised product range. This one-on-one communication takes place on your website or through your media channels.


The Recommender System is based on the powerful collaborative filtering technique, which allows us to determine the specific taste of every individual customer through the use of product engagement data. This data includes all touch points between each individual customer and an unique product. Based on this data, the system finds other customers with similar taste and marks these as lookalikes. Lookalikes are of crucial value because the same interest in products can lead to the same purchasing behaviour. So, based on data from these lookalikes, the system can select products they found interesting or that they have had some interaction with and then recommend these to customers.

One simplified example: a group of people who have been invited to a wedding are searching online for outfit inspiration. Wedding guest 1 looks like wedding guest 2 because both of them have shown the same interest in similar products. Therefore, the Recommender System considers them to be lookalikes. Guest 1 bought product X, so the system offers guest 2 product X as well.


All kinds of available engagement data between customer and product can be used to create personalised product recommendations. This includes productClick, productView (both long and short), addToCart, addToWishlist, transactions and other interactions that can occur between a product and a consumer. The Recommender System interprets interactions in a different matter, however, because using a different algorithm assigns unique weights to each unique type of interaction. The weight assigned to each type of interaction determines how much it affects the customer’s taste. Here, the purchase of the product is assigned the most weight and has a strong impact on the taste, while a productView is considered less important in the taste calculation.


When we built the personalised Recommender System, we took a number of common challenges into account. The fact that we incorporated these challenges adds significant value to the Recommender System. The challenges:

  • First of all, the cold-start issue – one of the main challenges. When new products are added to the catalogue or website, they are not automatically recommended to other customers. This makes sense, given that few to no customers have ever had an interaction with these new products. This is why we are not exactly sure for which customers and which kind of taste this product is relevant. We solve this by using product characteristics (brand, price, category, etc.). These product characteristics are used to match new products to existing products. With this information, we can determine the taste of all users in relation to the new product. And that’s not all; we can actually proactively push the cold-items when the recommendations are calculated in real time.
  • A second challenge is that sometimes it is not possible to recommend similar or additional products whatsoever for a specific product. This can be caused by a lack of statistical relevance, for instance. The solution for this is backfilling: recommending the most popular products instead.
  • A third challenge is that some products are subject to trends, like clothing. Meanwhile, other products, such as flowers, fruit and vegetables are extremely seasonal. We can include seasonality in the Recommender System, so we recommend products when they are ‘in.
  • The fourth point is not actually a challenge, but an addition to the system. We can tweak the algorithm in such a way that products with a large margin can be given higher priority. Please note: this is only done to the products that are relevant to the customer. After all, the goal is to personalise the products offering and enhance the customer experience.


It is essential to update the Recommender System frequently. This way, the system learns and includes new information into generating the recommendations. This is important to ensure that the recommendations to the customer are as accurate as possible. This learning process takes place in Azure ML (the machine learning stack from Microsoft), on which the entire system runs and where all data is stored. The learning process (also called the retraining cycle) comprises five steps, as shown in the figure below:


This training process is executed on a weekly or daily basis, depending on the type of business and turnover of new products, customers and traffic. As a result, new products and customers are included in the system and the system can also pick up trends and seasonal or other patterns extremely fast. In addition, the system can be completely integrated into the Azure stack of the business (if using Azure), though it can, of course, also be used as a stand-alone solution.


The Recommender System can be used on-site. This means that personalised alternative recommendations are given on the product page, the home page or on the category page, for instance, while additional/complementary products are suggested in the shopping cart. By using the system on-site, you both inspire customers and help them find the products they are looking for more quickly and easily. This enhances the user experience plus the average order value through cross- and upselling.

Recommender system
Recommender system

In addition to on-site use, the Recommender System can also be used in media. For example, it can be used to generate personalised e-mail content, offering alternative or complementary products. Or dynamic remarketing with personalised alternative or additional/complementary products based on the products the consumer showed an interest in during his or her website visit. So, we can also use these product recommendations in our dynamic retargeting campaigns and target customers with alternative products instead of just the products they viewed and did not buy, which would be the traditional dynamic retargeting method. This way, the customer is also inspired through advertising.


The Recommender System can be used for multiple industries, such as B2B, B2C, product-related services and more abstract services. The Recommender System can also be used to recommend services, FAQ or content, for instance. So, it is a very versatile system.


Thanks to the real-time character, the Recommender System is indispensable for website and media personalisation. So, what can you achieve with this? An enhanced user experience and more product sales. An enhanced user experience because the customer can find the right product faster and more easily and more product sales (cross-sell items) because the customers are alerted to matching, complementary products.

By including the ability to deal with multiple challenges such as cold-start, backfilling, seasonality and prioritising products with a higher margin during the construction of the system, it is a must-have for businesses that are focussing on personalisation and their online customer experience.

Questions? We are here to help!


If you're reading this, you unfortunately can't see the form that's supposed to be here. You probably have an ad blocker installed. Please switch off your adblocker in order to see this form.

Still encountering problems? Open this page in a different browser or get in touch with us: [email protected]