AI in action: product recommendations
Product recommendation is the process of identifying products that appeal to individual users and presenting them at an opportune moment as a way of encouraging cross-selling and upselling. For most ecommerce businesses, product recommendations have now long been an integral aspect of their operations, increasing revenue and user engagement.
Machine learning (ML) has allowed for the increased automation, accuracy, and timeliness of product recommendations, which in turn has produced massive increases in revenue for companies, particularly in the ecommerce sector. Salesforce found that “visits where the shopper clicked a recommendation comprise just 7 percent of visits, but drive an astounding 24 percent of orders and 26 percent of revenue.”
Super.AI has helped several large ecommerce sites to leverage their existing product data to begin providing targeted product recommendations to their users. In this post, we’re going to take a look over how product recommender systems work, what advantages they provide, and how super.AI can help you to leverage your existing data in a recommender system.
How does product recommendation work?
When making recommendations there are two key elements: users and products. From this pair, we derive three relationships:
- User-user relationships
- Users from similar demographics or with similar interests are more likely to be interested in similar products
- User-product relationships
- What users have previously purchased influences what they purchase in the future
- Product-product relationships
- Similar or related products are more likely to be bought together or around the same time
Recommender systems are built around using these three relationships to filter data and arrive at an ideal product selection. There are three primary methods of filtering data to achieve this:
- What products have similar attributes to a recently viewed or purchased product?
- What did similar users buy who also viewed or purchased a product?
- Complementary (or cluster)
- What products go well together, regardless of user behavior (e.g., smartphone and screen protector)
- A combination of collaborative and content-based filtering
Benefits of product recommendation
As ecommerce becomes an increasingly competitive space, maximizing efficiency and finding effective cost-cutting techniques has become essential to survive. Exposing users to timely, personalized purchasing suggestions means users are more likely to view and purchase products they might not otherwise have seen. Monetate found that when users engage with product recommendations, it “can lead to a 70 percent increase in purchase rates, both in that initial session and in return sessions, and 33 percent higher average order values.” As far back as 2013, McKinsey found that “35 percent of what consumers purchase on Amazon and 75 percent of what they watch on Netflix come from [algorithmic] product recommendations”. If you have access to user reviews or recommendations, you stand to gain even further. McKinsey again: “peer recommendations carry ten times more weight than recommendations from salespeople.” This opens onto trust-based recommender systems, which harness online peer groups as an additional form of data filtering.
How super.AI makes product recommendations work
The single most important element in building an effective recommender system is data. You need to collect, store, and organize as much data on your user-product relationships as possible. The more data you have, the more information your recommender system has to learn from. At super.AI, we’ve worked with two global ecommerce brands that have adopted third-party vendor marketplaces in an effort to expand profit pools.
If you run an ecommerce site with third-party vendors and user-generated content, there’s a wealth of data already at your fingertips. The essential first step to successfully harnessing the data is two-fold:
- Create a product taxonomy
- Categorize user behavior
We help with the former. Creating a taxonomy is an enormously challenging task in any scenario, as there is always plenty of room for ambiguity. This is never more true than when your product database is made up of user-generated content. In this case, it’s not always easy to tell what a product is, let alone where it should fit into a taxonomy. The solution we developed involved the use of our image categorization and image tagging data programs to automatically add metadata to hundreds of thousands of images from their marketplaces. We identified key information about each product and built a comprehensive product profile of tens of thousands of products. All our clients had to do was send us their raw data. The process on our side looked like this:
- Our image categorization data program broke up the raw input data into discrete tasks, applying a category to each data point, thereby creating a product taxonomy
- We ran our image tagging data program to add metadata to the images within this taxonomy, returning labeled training data while simultaneously training a model to label any additional input data we received
Here’s just some of the information or metadata we identified for each product:
- Type (dining table, sofa, bed, etc.)
- Where the image was taken (studio, home, etc.)
- Whether the image showed the product in full or was a close up
- If the image showed the product’s size
2) Pet supplies
- Size or volume
- Applicable dog breed
- Type (baby, collectible, fashion, etc.)
- Whether the product is still in its original packaging
- If the image showed the product’s size
Once we had produced the product taxonomy and applied metadata to the images, the hard work was complete. Our clients were free to use the newly created training data as they needed. As an optional final step, super.AI can also use this labeled data alongside purchasing data to train a recommender model, which would be accessible via API. If you think your ecommerce site could benefit from a recommender system, reach out to one of our experts and we’ll work to produce a solution tailored to your data.