ML Is Redefining Product Recommendations for eCommerce

ML (Machine Learning) and AI (Artificial Intelligence) tools are becoming more and more affordable and approachable. The complexity and cost thresholds a problem has to touch to qualify for an ML treatment quickly go down. Many small tasks, such as providing product recommendations to eCommerce website visitors, can now be approached with ML solutions. The benefits are significant.

For years the big eCommerce players have leveraged ML to provide individualized, dynamic, and practical recommendations. But, unfortunately, small and medium eCommerce players didn’t have the budget, resources, and volume to integrate ML solutions into their systems. But now, the landscape is changing. New companies are coming to the market proposing affordable ML solutions that can be quickly and effectively integrated with almost any eCommerce platform.

Old Product Recommendation Systems

A classic way to generate product recommendations is to analyze the collection of converted carts (orders) and use an algorithm to create a prediction about the product a customer is more likely to purchase. When the number of orders is too small, companies can include in the analysis also the abandoned carts. One of the most popular algorithms to analyze the recent collective purchase history of the customers of an eCommerce website is the following:

To find the best recommendation for a user who purchased or added to the cart product P, the most straightforward contextual recommendation algorithm can be summarized with the following steps:

  1. Select all recent orders with product P and at least one additional product
  2. Create a list of products appearing in orders with P, and for each, calculate the frequency.
  3. Sort the list by frequency and get the top product.
  4. If the top product doesn’t violate any business rule, you are done, otherwise, discard the top product and repeat this step with the product in the following position.

Business rules are usually specific for each industry. For example, if an eCommerce store sells clothes, the product to recommend on product P should be for the same gender and possibly for the same season. If, for example, P is a bundle, the recommended product should not be a component of the bundle.

New Product Recommendation Systems

One of the most popular ML recommendation algorithms, based exclusively on the purchase history, is Collaborative Filtering (CF). Collaborative Filtering takes into account both similarities among products as well as similarities among users. For a complete description of the Collaborative Filtering algorithm, I recommend the Google pages on Recommendation Systems.

A more advanced Recommendation System can also consider additional signals, like the navigation history of the user, his profile including preferences, body size and type (for clothing stores), and current sales trends.

On top of that, the recommendation system has to filter the recommendations based on business rules. As in the classical recommendation engines, business rules can vary from industry to industry and even from one product category to another. Specific business rules should also drive the inclusion or exclusion of products currently undergoing some critical promotion. Some of these promotions could be limited to customers meeting certain conditions that the recommendation engine can not quickly and easily check. In the past, I had to deal with some crazy promotions from manufacturers like “only for people who didn’t purchase anything in the last two years” or “only for owners of a certain product who didn’t upgrade in the last X days.”

A dynamic recommendation engine should also be connected to the stock management system to avoid recommending products out of stock. Finally, a good approach should be to avoid recommending products that received a bad review, at least until they get a new, more recent positive review.

What are the benefits of an ML Recommendation System?

The use of ML is not enough to guarantee superior quality for the recommendations. However, ML systems leveraging multiple signals, such as stock level, product reviews, trends, navigation history, and industry-specific business rules, can create top quality and hyper-personalized recommendations to generate high conversion rates. Here is a partial list of the benefits of an ideal ML recommendation system:

  • Personalization: Each visitor gets a different personalized recommendation based on their profile, purchase history, visits history, etc. E.g., after purchasing an item, customers receive recommendations significantly different from what they were getting before.
  • Dynamicity: Recommendations are continually updated with the latest and most updated information available. The recommendation engine should gather signals from user history, visitors’ behavior, market status, stock level, etc.
  • Transparency: If a model, a color, or size is missing, the alternative recommendation comes with an an explanation for the customer explaining the compromise.
  • Relevance: The recommendation is always relevant and helping the customer. It fits in their customer journey. It’s not, and it doesn’t look like pitching a sale. The user perceives being helped, not being pushed.
  • Actionability: The recommended products can be purchased immediately. They exist, are in stock, and are available in the size, color, and options matching the visitor’s profile. If a recommendation includes a product in a promotion, the recommendation engine already checked the visitor qualifies for the promotion.
  • Usefulness: The recommended product is not just a variation of what the user already purchased or added to the cart. It is the product most likely to be bought by that user at that moment. E.g., if a user just purchased a smartphone, the recommendation should recommend headsets, chargers, phone covers, etc., not another phone.
  • Variety: Recommending always the same product or only the most well-known items can make the customers bored. The recommendation system should add some variety. The major obstacle to variety is sparsity. For the less popular products, there could not be enough data to make a recommendation confidently. The risk is that these products could become part of a self-perpetuating cycle and never be included in any recommendation.

What are the options on the market for an ML Recommendation System?

I didn’t have the option to test drive any of the options available on the market, but there are a few products to consider:

Franco lives and works in the eCommerce territory, a wild area between the Kingdom of Technology and the Land of Marketing. He speaks fluently the language of both realms. For many years, Franco has been helping people bridge the divide and successfully collaborate.

If you want to find out more about Franco, visit his LinkedIn profile or send him an email folini[at]

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s