Product Recommendation Models for E-Commerce Success

As one of the first ways to provide online personalization, the product recommendation model changed the way e-commerce worked. Anonymously browsing through generic stock for an elusive item became a thing of the past since Amazon filed its first patent for the Amazon recommendation engine in 2001. Today, up to 45 different recommendation widgets are visible on the Amazon app homepage alone. Small wonder that 35 percent of what consumers purchase on Amazon comes from product recommendations, according to McKinsey.

Serving the Long Tail

Recommendation models help customers to narrow down choices and are now a solid part of the web experience. Whenever you use services like Netflix, YouTube, LinkedIn or Facebook, you will be offered personalized suggestions on movies, videos, jobs or pages you might like. The root of recommendation models lies in the unlimited space that the internet has given us. In physical stores, there is a limited shelf space to store goods, which results in shopkeepers only displaying their most popular items. This serves as a pre-filter.

With the rise of e-commerce, there suddenly was space and opportunity to offer the less popular products. Chris Anderson, the editor in chief of Wired magazine, dubbed this the Long Tail Effect. He predicted that because everything is becoming available to everyone in the Digital Age, niches will rise. Consumers are increasingly shifting away from a smaller number of popular products, which make up the head of the demand curve. Given enough traffic, products with a low sales volume can collectively make up a market larger than that of the most popular, best-selling items.

The Impact of Product Recommendation Models

In his book The Long Tail, Anderson illustrates the power of product recommendation models with a story about the book Touching the Void. This mountain-climbing tragedy was written in 1988 and after modest success, soon was forgotten. Then, a similar book called Into Thin Air became a publishing sensation a decade later. Suddenly, the first book also started selling again, even outperforming the second book.

The benefits of using product recommendation models are manifold. Here are three key figures to remember:

• 90% of consumers find personalization appealing. (Epsilon)
• Consumers that clicked on recommendations are 4.5x more likely to add these items to their cart and have a 5x higher per-visit spend. (Salesforce)
• Product recommendations account for up to 31% of e-commerce revenue. (Barilliance Research)

Finding the Right Approach to Product Recommendation

The prime goal of product recommendation is to display only meaningful items to the customer to increase the sales of the company. At the beginning of e-commerce, there was one main way to do product recommendation online called collaborative filtering. This system analyzes and identifies similarities to serve customers recommendations. These similarities can be between users and item interactions:

• Item-item collaborative filtering: recommendations are given based on customers’ past purchases or product ratings. If you rate a certain product 10/10, then you will be shown products with similar attributes. Here, the online retailer’s aim is to produce a finite list of the best things to recommend to a certain customer.
• User-user collaborative filtering: recommendations are given based on the preferences of other customers with similar purchase histories. This is about predicting the rating value for all customer-item combinations.

From the start, Amazon decided to choose item-item collaborative filtering because inspecting recent-purchase histories would mean far fewer lookups than identifying similar customers. The crucial question Amazon had to answer was how to measure relatedness. Counting how often consumers who bought product A would also buy product B wouldn’t work: bestselling books and trash bags would end up as the top recommendations. Amazon decided on using a relatedness metric based on differential probabilities: product B is related to product A if purchasers of A are more likely to buy B than the average Amazon customer is. The greater the difference in probability, the greater the items’ relatedness. Amazon researchers published their findings in a 2003 paper called “Amazon.com Recommendations: Item-to-Item Collaborative Filtering”, which was identified in 2013 by the journal IEEE Internet Computing to have best withstood the test of time.

What Makes Product Recommendation Work?

To make product recommendation models successful, there are four desired goals:

• Relevance: Recommendations make sense only when they relate specifically to the user. Items that customers find interesting are more likely to be purchased or consumed.
• Novelty: A recommendation is more likely to be useful if it appears to be something the user has not seen or consumed before.
• Serendipity: It is also possible to boost sales by recommending items that are unexpected. However, serendipity is not the same as novelty. Serendipity can be defined as making happy and unexpected discoveries by accident.
• Diversity: It is also important to increase the diversity of recommendations. Recommending items that are similar isn't very helpful.

Today, collaborative filtering is one of many approaches, now often used as part of hybrid models. Other noteworthy approaches are:

• Content-based filtering: based on a description of the item and a profile of the user's preferences. Best suited to situations where there is known data on an item (name, location, description, etc.), but not on the user.
• Session-based filtering: This approach uses the interactions of a user within a session to generate recommendations. YouTube is a typical example of this approach.
• Risk-aware filtering: Most approaches do not consider the risk of disturbing the user with unwanted notifications. By incorporating this risk into the recommendation process, the performance increases.

Amazon reports a 2x typical algorithmic improvement after trying out various approaches and finally found a new model for recommendation using multilayer neural networks for classification, which trains a model using a loss function. Jeff Wilke, consumer division CEO explains in this short video the history of Amazon’s recommendation algorithm.

It is really important to understand what your consumers want and offer personalized, authentic experiences. With Multidomain Master Data Management you can create a view of your consumers needs, to then build an advanced personalized strategy that helps you shape big successes. 

Making the Best Product Recommendation Model Possible

At the core, a product recommendation model is about enhancing the customer online shopping experience through personalization. This is made possible by understanding the customer’s history, preferences, behavior and context. But for an optimal experience, that’s not all. Recommendations for out-of-stock products would be counterproductive. You might want to push certain products which go out of season soon, or that offer you a higher margin. You can bear in mind the richness of the product information available, and only recommend products that are more likely to sell. And if you have several stores and offer omnichannel services like same-day delivery and pickup, you will want to consider product availability per store.

Amazon founder Jeff Bezos famously said: “You collect as much data as you can. You immerse yourself in that data. But then make the decision with your heart.” To make the best product recommendation model for your organization, you need to include four data domains: customer data, product data (PIM), location data and supplier data. All the processes, roles, policies, standards, and metrics needed to ensure the effective and efficient use of data within your organization is called Data Governance. If you get this right, your product recommendation will stand out in the market.

This is the fourth blog of our article series on Data Governance. Click here for the next article.