The Recommendation Algorithm
Learn how Amazon designs product recommendations using a specialized algorithmic system based on a variety of data to provide tailored suggestions.
How does the algorithm work?
Product recommendations are a central component of eCommerce to offer customers suitable recommendations that are optimally tailored to their needs. Amazon uses its own algorithmic recommendation system for this, which accesses various data sources. Product data is particularly important, especially when introducing new products. This product data plays a crucial role in how well the algorithm can integrate new products into its recommendations, which is examined in more detail below.
Algorithm and recommendations
Amazon's recommendation algorithm is based on three main building blocks: First, it uses data on the purchase history of a large number of customers to identify products that are frequently purchased together. Second, data on common views of different products for the same search query is analyzed to recommend similar products. Third, the algorithm includes product features to suggest new relevant products that have rarely been purchased or viewed so far.
How does the algorithm work?
The recommendation algorithm works with a product graph in which nodes (e.g. A, B, C) represent different products. Each node is assigned metadata such as product properties. The arrow lines between the nodes represent either common purchases (yellow) or common views (red) of the products. This product graph forms the basis for training a neural network (dark blue), which ultimately generates accurate product recommendations.
Product networking
In product networking, a distinction is made between unidirectional and bidirectional connections. In a unidirectional connection, the purchase of one product depends on another, for example, the purchase of a smartphone case depends on the purchase of a smartphone. Here, the smartphone is the source and the case is the target. A bidirectional connection, on the other hand, exists without dependency between the products, such as when a smartphone case is purchased in connection with a display protection film. If two similar products are frequently viewed together, the connection is also bidirectional.
How can relevant products be found?
For each product, two environments are created: In one, the product is considered as the source of a product recommendation, in the other as the target. This structure serves to avoid asymmetries. For example, if a smartphone case (Product A) serves as the source for product recommendations, no smartphones (C, B) should be suggested as target products, but more suitable accessories such as display protection films (E).
Conclusion
When a customer purchases a smartphone, a target environment with new, relevant products can be created on this basis, which includes, among other things, frequently purchased products together. By recommending products that are often viewed together, a greater variety of similar products is presented. The inclusion of metadata (relevant product properties) ensures that new products are not overlooked. For the algorithm, it is crucial that comprehensive metadata is filled out in addition to a detailed product description in order to be able to recommend the product effectively.
