Algorithms and Amazon 🤖💻

“Because of their opacity, algorithms can privilege or discriminate without their creators designing them to do so, or even being aware of it.”

The Atlantic: Algorithms Are People

What’s an Algorithm?

An algorithm is a set of instructions that a computer follows while completing a task or solving a problem. The instructions can be represented through computer programs coded by engineers. The problem can be as trivial as adding two numbers, or as complex as figuring out what products to show to which customers.

At a company with services and offerings as myriad as Amazon, algorithms may be used in innumerable ways – from routing delivery vans optimally or detecting what a user is saying to his smart home system, to figuring out what recommendations and rankings to show to a particular consumer online. Amazon even uses algorithms to monitor and discipline warehouse workers!

Cool. But why are they important? Examining under-the-hood systems like algorithms is important because they directly control the content that users see, influence how they interact with it, and dictate how information and data about users are handled. When these services become as socially pervasive as Amazon is, these systems can have profound effects at large scales.

Image from A9 – one of the Amazon subsidiaries that helps build the product search and ranking algorithm.

The Consumer’s Case

A common type of algorithm that Amazon employs for its online store is a recommendation algorithm. This is responsible for personalizing homepages, catalogs, search results, etc. for different users – providing the ‘best’ recommendation for the consumer. One of the techniques Amazon has used for this is called collaborative filtering – ‘collaborative’ because it models a user’s preferences based on other users, or an item’s relevance based on other items. The algorithm has over the years expanded to factor in “personal preferences such as brands or fashion styles” and learned to “time recommendations (you may want to order more diapers!)”, as mentioned on an Amazon science blog.

The reservoir of user data that feeds an algorithm like this can be very deep. For example, to learn about a user’s preference, the algorithm can look at:

Previously purchased items, and their ratings
Items added to carts or wishlists but not bought
How long you dwell on a product page before buying it
Click-through data (the path you took through the website and different products)
What time you usually purchase a particular product

Because of the intricacies of how the algorithm makes decisions and suggestions, it is often difficult to assign accountability when something goes south (for example when reports showed how Amazon’s algorithm recommended that customers buy items that can be used to create explosives!).

In another report from 2016, African-American and Latino customers were shown to be disadvantaged when their minority urban neighborhoods were ineligible for Prime same-day delivery. This was because the Prime algorithm takes into account information like distance from warehouses, the number of Prime members in the area, etc. but neglects the fact that race and poverty correlate with housing – a correlation that is thus mirrored in Amazon’s offerings to its customers!

The fact that algorithms are built by humans, and learn from humans means that societal biases creep into the outputs and decisions made by the algorithms. Because of their opacity and technical complexity, it is also hard to setup a system of accountability.

The Seller’s Story

How are products and sellers ranked? How are some sellers promoted over others?

Amazon’s subsidiary A9.com is the algorithm responsible for the ranking of products and sellers. Under this algorithm, seven key traits are accounted for when ranking a product: Relevancy to searched keywords, pricing, conversions, product listing completeness, stock supply, sales rank, and reviews. Matching a product to searched keywords is necessary to promote a product because the algorithm wants to provide the most accurate product to the consumer searching for the product. Competitively and accurately pricing a product is necessary to become a higher ranked product; overpriced products will not sell, and underpriced products might appear suspicious. Higher conversion rates of a product are important because they will ensure that Amazon shoppers will return to the product. Being thorough and completing a product listing is important because it provides the consumer with all information needed to be more likely to purchase the product. Stock supply is essential because the algorithm will not place items that are constantly out of stock in a high-ranking position. The sales rank of a seller is determined by comparisons between similar products and the number of conversions, more conversions equals a higher ranking. Arguably the most important aspect of a product is the reviews a seller receives. The better the reviews a seller receives, the higher the product will be ranked.

Amazon choice/Amazon basics products ranked at the top – algorithmic monopoly?

Amazon’s basics/choice line also dominates a variety of different products being sold on the website. The example provided shows two similar sets of nonstick pans. While they look similar, the Amazon basics product is being sold at a fraction of the cost. This disincentivizes the consumer to purchase a product of higher quality because of how inexpensive the Amazon basics product is compared to the 3rd party seller’s product. With more sales being funneled through the Amazon basics/choice products, the algorithm recognizes the product being sold more and ranks the product higher.

Pricing, seller reviews. Sponsored products.

Purchasing a sponsored product plan gives sellers an advantage over other products, even if the product itself is not as popular as other products. When purchasing a sponsored product plan the seller can attach certain keywords to the product being sold and have that product pushed to the top of the search results. The product may not be causally related to what the consumer is looking to purchase, but by being at the top of the list consumers are immediately drawn to the top product. There is a huge advantage to being at the top of the page because around 70% of consumers never click on the second page of the search results of the product they are looking for.

What’s the fuss?

Reinforcing Stereotypes

When looking up “toys for girls” versus “toys for boys”, we can see that the results that come up are inherently based on the societal expectations and ideals we have of what boys and girls should play with. These algorithms further reinforce and perpetuate stereotypes that boys have to be hyper-masculine and only play with products with dark hues/colors and for girls to play with pink-colored items that are safe and promote being submissive and feminine. In recognizing this, however, we also need to take a look at how the algorithm works as it simply enhances and promotes what has been successful in the past. This means that best-selling toys for boys tend to fall under a specific category and those for girls do the same.

Project Recognition

Project Rekognition was Amazon’s facial recognition software developed in 2016 to be used in their suite of products. A major finding was that the software was reported to have trouble correctly identifying darker complexion individuals compared to their lighter skinned counterparts. This, coupled with the fact that Amazon supplies US government agencies such as ICE with this technology to be used on the job, is a major negative consequence of these algorithms.

References

Algorithms Are People, 2019. Sidney Fussell, The Atlantic.
Amazon Knows What You Buy…, 2019. Karen Weise, The New York Times.
Researchers find gender and racial bias in Amazon’s facial recognition software, 2019. Nicole Karlis, Salon.