Skip to main content



Amazon.com recommendations and Hub-Authority Networks

The article linked below titled “Amazon’s Recommendation Secret” claims that Amazon.com’s (Amazon) increase in growth is due to its successful integration of recommendations on its website. To summarize their recommendation method when looking at their site: a typical user could search for a certain keyword and then a list of items will show up. In addition, when an item is clicked, on that site there will be another list of items at the bottom of page under “frequently bought together” or “sponsored products related to this item”. The article lists that Amazon tracks metrics such as “open rate” and “click rate”. In addition, their hidden secret is that they also try to provide recommendations that would encourage the highest revenue.

For a more thorough explanation, the document titled “Amazon.com Recommendations” (http://www.cs.umd.edu/~samir/498/Amazon-Recommendations.pdf) actually delves extensively into the specific recommendation algorithm Amazon uses. It is called “Item-to-Item Collaborative Filtering”. It highlights that Amazon doesn’t use the typical method that groups customers by their purchase and click patterns and then subsequently predicting what the customer might want to purchase. Instead, they focus on the similarity of items being purchased and then create a matrix of data to find other items that are similar based off of past purchases. That matrix includes different vectors of information between items, including the item’s rating, price, and popularity.

This process can be simplified to represent hub-authority network (obviously it is much more complex than this). You can think of a click or purchase as an edge pointing from an item to another item. And like the document argues, the algorithm focuses more on item similarity rather than user action similarity, so the nodes would be items to purchase rather than customers. So, with this network, the higher the clicks to an item, the higher the normalized authority score. And then with these scores, Amazon uses other computation methods to provide influence over ratings and potential profit in the final recommendation list. They include methods such as calculating the angles between edges to judge similarity between items. After all this computation, then Amazon has a recommendation list that is similar to items you purchased, thus leading to a higher potential profit.

http://fortune.com/2012/07/30/amazons-recommendation-secret/

Comments

Leave a Reply

Blogging Calendar

October 2018
M T W T F S S
« Sep   Nov »
1234567
891011121314
15161718192021
22232425262728
293031  

Archives