Collaborative filtering: an ace up Cartamundi’s sleeve

Jun 18, 2019

operations
IT
discrete manufacturing
professional services

‘The brand you don’t know you love’: that’s how Belgian card manufacturer Cartamundi introduces itself on its website. And if you’ve ever played a parlor game, there’s a big chance you’ve been dealt cards produced by Cartamundi. When it comes to operational efficiency however, they are not playing around. To optimize the way its employees are using SAP, Cartamundi is experimenting with a technology called ‘collaborative filtering’.

Challenge: Cartamundi wanted to harvest the full potential of implemented SAP solutions. A recommendation engine could guide users to SAP transactions that they were not familiar with.
Solution: Map all SAP transactions within the company: which modules are used together or by similar profiles? Based on this data, the recommendation system can propose new functionalities to maximize the user's efficiency.
Result: Insights that help optimize operational efficiency and lay the groundwork for other experiments with AI. Improved SAP usage also leads to more value for the same licences.

Much like Cartamundi’s cards, collaborative filtering is one of those things you didn’t know you already knew. “Every time you get customized recommendations from Netflix, Spotify or Amazon, that’s collaborative filtering at work,” explains Katrien Sterken, SAP consultant at delaware. “It’s the engine behind the personalized web experience we have come to appreciate so much.”

Collaborative filtering for business optimization

While customization of recommendations is a major revenue driver for countless businesses, the applications of this technology go beyond finding new tunes to listen to or series to watch. As part of a DEL20 experiment, Cartamundi decided to investigate how collaborative filtering can play a role in identifying how SAP is used throughout the business.

The goal is twofold:

To guide users – through recommendations – to SAP transactions that they may not have heard of before, but that are likely to help them perform their jobs in a more efficient way.
To obtain industry insights regarding the usage of SAP transactions and modules by comparing the data sets of multiple companies.

This, of course, required careful mapping of all SAP transactions within the company:

Which modules are most used together?
Which modules are used by similar profiles?

To gain more complete insights into the use of SAP, transactional data from other companies willing to cooperate with the experiment was incorporated as well. This made it possible to benchmark against similar setups in terms of which transactions are most used by similar users across companies.

“Based on these insights, the recommendation system is able to ‘fill in the blanks’ in terms of what transactions are likely to be useful for a certain user and will help maximize the user’s efficiency,” explains Katrien. “A great example was a production planner who looked at every sales order in SAP individually through transaction VA03. This person didn’t know there was also a transaction VA05N, which is a list of sales orders report with various selection criteria. This handy report shows all the needed sales order data in one view.”

delaware’s DEL20 program provided us the opportunity to further develop ideas like the SAPFlix experiment with their technology experts in a conceptual phase. It’s a win-win for both companies.

Mark Van Den Bosch, CIO at Cartamundi

The limits of collaborative filtering

One of the biggest challenges the experimenters were soon facing was the size of the data set. “We had to limit our extractions to 3 months to keep processing feasible,” explains Katrien. “That’s why we’re currently looking to switch to a machine-learning model that will allow us to process even more data and make better predictions.”

Another hurdle is context – something most recommendation systems today are still lacking. “How many times have you been stalked on Facebook by a vintage watch or another item that you’ve already bought? ‘Errors’ like this show that context is still a major gap today. In our Cartamundi experiment, this manifested itself in users getting recommendations for modules they didn’t have access to. In the next phase of the experiment, we will be adding context to make recommendations even more accurate and relevant.”

How collaborative filtering works

Recommendation systems use collaborative filtering to identify and predict behavioral patterns, e.g. what we ‘like’, buy, watch, listen to, etc. In ‘item-based collaborative filtering’, you compare two items each time and determine the overlap between two users for these items. The bigger the overlap, the stronger the similarities between both items. In ‘user-based collaborative filtering’, you determine the overlap in items between two users. This gives you an idea of how comparable the users are. To make relevant recommendations, the system identifies the most similar user and takes an item the initial user doesn’t already have.

For example, if users A and B both like almost all the same series on Netflix, but user B has seen (and liked) Stranger Things and user A hasn’t yet, Netflix will likely recommend this series to user A. To be able to do this, collaborative filtering requires a large number of active users to make accurate predictions and/or recommendations. The more data you have, the better the recommendations will be.

We can predict whether user E will want the headphones by comparing his tastes with those of users who have indicated similar tastes in other items. In this case, the distance between users B and E is the same as the distance between users C and E: both have 2 items overlap. Since B and C are the most similar users, and both don’t want the headphone, there’s a good chance E won’t want the headphone either.