Turning multi–dimensional transactional data into longer term revenue stream undoubtedly requires advanced analytics that goes beyond available BI toolset.
In 2024, an online mobile gadget store approached us to explore the possibility of running a promotional campaign via email. The e-commerce data available was split into twofold: previously purchased items and what they call “orphan” shopping cart items.
Unlike previous campaigns, the client wanted to combine existing e-commerce data with incomplete profile information on their end customers. The new recommender system was to generate custom item suggestions for each shopper on the target list and would run in the back office of the client’s marketing department.
All data was anonymized during implementation and production.
Client ... tight spot
The client’s marketing department had searched the AI field for a recommendation system that can combine transactional information with profile data. The search was without success, since most of what is available (which is quite advanced) relies on affinity or collaborative filtering – ML algorithms which are based on two simple assumptions:
- Shoppers are inclined to like things similarly related to other things they like, or
- Shoppers may like things that are chosen by other people who already choose items similar to the ones they bought or liked.
Even though the mathematics behind affinity and collaborative filtering is quite advanced, its main problem is that it solely relies on item data. The client wanted to put in use some of its “treasured profile” data to engage end customers and re-target them with “higher precision” for greater conversion rates.
GoFlek ... to rescue
For GoFlek this was a great trial opportunity to run its novel Recommendation Engine firsthand since it had just added this capability to the Flek Machine and was ready for the new use case. Moreover, it was a great occasion to test Flek’s serverless mode which allows running its core probabilistic engine on the cloud without the need for a dedicated server – an essential condition set by the client to increase scaling and save on project cost by paying only for what is used.
Flek Recommender … how it works
Recommendation in Flek utilizes probabilistic algorithms that make it easy to combine both profile and item based techniques at the same time.
Internally, the recommendation system runs a separate ranking algorithm for each of the 2 domains: ITEM and TRAIT. It then combines the 2 rankings into one aggregated ranking using a Generalized Weighted Merge (GWM) algorithm. All 2 domain rankings plus GWM algorithms were developed by GoFlek engineering and optimized for rapid recommendation.
To glue everything together and interface with the client’s email engine, GoFlek consultants used the Flek Toolkit and APIs for data preparation and encoding, ML model generation and recommendation execution.
The main challenge was the incomplete profile records available. This required explicitly feeding the recommender system different types of input or missing information on each end-customer. For instance, data regarding gender, age, or a combination of both was utilized based on the information available in the respective records. Of course, the input information was much more complex and custom coding was improvised to run the whole pipeline smoothly.
The entire project took about 2 weeks to complete and then 1 week for testing and another for putting in production which was right in time and about 1 month before Black Friday.
