Personalized cross sell recommendations

cross sell

Does cross-selling sound pushy and repugnant? It does somewhat, especially when a brute force, one-size-fits-all approach is deployed to overawe your customer. In contrast, how does it sound that you are making sure your customer has everything he needs to use your product? The concept is identical; you make a targeted intervention, to help your customer to add articles to his shopping basket. However, aiming for a better customer experience will improve basket size as a result, while only aiming for higher basket size may deteriorate other important metrics.

So, how can you have product recommendations that contribute to a great overall experience of your customer? Be relevant. 

In general, relevance makes or breaks the customers experience. Do recommendations fit the purpose of the shopping trip of your customer? When a customer is looking to buy an action camera, he may like recommendations for waterproof casings. Context is also important. While your customer is on the product page of an action camera, he may prefer recommendations of various similar action cameras, while the shopping cart page may remind your customer of the waterproof case or necessary batteries. Lastly, familiarity: products or brands that commend the interest of your customer are more likely to be seen as relevant.

Saint-Gobain, who sells construction materials to building companies, recognized how cross-sell recommendations could help customers ordering all the essential materials needed for a job and thereby avoid loss of productive time associated with a second trip for picking up missing items. Furthermore, in their industry a customer, e.g. a carpenter, often has strong preferences with regard to the materials he works with. In order to achieve the desired quality, productivity or reliability, he relies on familiar materials, brands or models. Personalized recommendations were therefore needed, generated by a system that could handle a complex set of requirements.

Relevance may mean familiarity with the product, the brand or the model, it depends on your business 

Saint-Gobain selected SPARQUE to get the solution they needed. Using SPARQUE, recommendations can easily be modeled to reflect your exact requirements. For Saint-Gobain, a cross-sell recommendation should first of all reflect what a particular customer had bought before. When plugs are recommended to go with plasterboards, the plugs he had bought before with plasterboards get a boost. Of the items that fit well as a cross-sell, but had not been bought before by the customer, the brands that are familiar get a boost. If he purchased a lot of materials of brand “Fisher”, the plugs from Fisher will be recommended, even though he has not bought any Fisher plugs before.

Personalized product recommendations is one example of real-world challenges in which eCommerce managers want data from various systems combined to produce relevant results for the end-user. A distinct advantage of SPARQUE is its flexible data model that is expressed as a graph – Google and Facebook use similar graph models. 

Using SPARQUE, all original data structures and relationships can therefore be preserved and used. That has great advantages, for an example, a B2B wholesaler has 100K+ articles and many items are in the long tail with few, if any, sales to generate cross-sell recommendations from. Using product properties stored in their PIM, we could model what were similar products and use those recommendations to overcome this cold start problem. The effort you invested in master data can now really be levered to generating relevant cross-sell recommendations.

Do you need a lot of data to do personalized recommendations? Most likely, you already have the data you need. 

In the use case above, product properties such as brand comes from the company PIM, individual and collective purchase history comes from the ERP system while user-id, items currently in the shopping cart and pages visited in the current session are stored in the session cookie and sent to SPARQUE as part of the query parameters.

Transaction data forms the core for the recommendations. The relevance of purchase history of the customer is high if the purchase has been done recently and decreases for purchases further in the past. Products that are frequently found together in orders provided the basic signal for what is a good cross-sell recommendation. For example, 6mm fixings will match well to 6mm plugs, as this combination is more commonly found in the purchase history than 6mm fixings with 8mm plugs. Using something that is similar to TF-IDF, the more unique certain combinations are, the stronger the signal. Recommendations are not limited to pairs either; all products that are already in the cart play a role. The better a product matches the combination of products in the cart, the higher the relevance score. With more items in the cart, the precision increases of what items match well and are therefore good cross-sell recommendations.