Case study: Amazon's Recommendation System #23
Conversation
First draft; text only.
|
|
|
No broken urls! 🎉 |
|
Re-rendered previews from the latest commit:
Updated at 2023-12-20 with changes from e01978f |
added citations for Amazon's rec system
Added slides
There was a problem hiding this comment.
Looks great, love the images.
There are a few dates and such missing from the citations - make sure these look okay when you check out the preview.
I was also thinking the case studies for course 1 should be moved to the end after 01f-AI_Possibilities-ground_rules.Rmd and could even be their own chapter "Case Studies". That way, you don't have to worry about defining terms ("neural network" "encoder" etc).
Make sure you add any spelling flags to the dictionary.txt file if they are actually words and you don't want them to be flagged by the spellchecker.
|
|
||
| In the early 2000’s, Amazon’s recommendation system was based on a method called “collaborative filtering”. The term comes from the fact that the method tries to predict a customer’s buying habits based on other similar customers in the marketplace. | ||
|
|
||
| Here is how it worked: The history of a customer on all items in the marketplace is stored as a large numerical list. Given a customer’s purchasing history list, the collaborative filtering algorithm looks for similar customers with similar purchasing patterns via a similarity score. The algorithm makes product recommendations based on their these similar customers’ top purchases. This is called customer-based collaborative filtering. |
There was a problem hiding this comment.
How about breaking this paragraph down into bullets or numbers?
|
|
||
| Here is how it worked: The history of a customer on all items in the marketplace is stored as a large numerical list. Given a customer’s purchasing history list, the collaborative filtering algorithm looks for similar customers with similar purchasing patterns via a similarity score. The algorithm makes product recommendations based on their these similar customers’ top purchases. This is called customer-based collaborative filtering. | ||
|
|
||
| However, this recommendation process was not particularly efficient: computing the similarity score between all customers was slow, and this had to be done regularly, as customers purchasing habits changed regularly. The system was not very stable. @sarwar2001item |
There was a problem hiding this comment.
I'd put the citations throughout within brackets, eg.
@sarwar2001item>[@sarwar2001item]
|
|
||
| As the number of Amazon users grew faster than the number of marketplace products, scientists thought about creating more stability by looking at the purchasing history of a product by customers instead. This is called item-based collaborative filtering. Given a customer’s recently purchased items, this new algorithm finds items that have a similar purchasing pattern, and recommend it to the customer. | ||
|
|
||
| A product’s purchasing history by all customers did not change quickly, so computing the similarity between all products did not have to be done regularly. This system was much more stable. @amazon-recommendation-history |
There was a problem hiding this comment.
Date missing on this ref?
|
|
||
| There are still many active areas of research to improve recommendation systems! | ||
|
|
||
| For instance, the order of purchase should be considered. A customer is likely to buy a smartphone, then buy a phone case, but not the other way around. Scientists have started to explore modeling items using a network, and using arrows between items to represent the likely purchase order. @amazon-gnn-recommendation |
There was a problem hiding this comment.
date missing on @amazon-gnn-recommendation and @linden2003two
|
No quiz formatting errors! 🎉 |
First draft; text only.
Purpose/implementation Section
What changes are being implemented in this Pull Request?
What was your approach?
What GitHub issue does your pull request address?
Tell potential reviewers what kind of feedback you are soliciting.
New Content Checklist
New content/chapter is in an Rmd file with this kind of format and headers.
New content/chapter contains learning objectives.
Bookdown successfully re-renders and any new content files have been added to the _bookdown.yml.
Spell check runs successfully).
Any newly necessary packages that are needed have been added to the Dockerfile and image.
Images are in the correct format for rendering.
Every new image has alt text and is in a Google Slide.
Each slide is described in the notes of the slide so learners relying on a screen reader can access the content. See https://lastcallmedia.com/blog/accessible-comics for more guidance on this.
The color palette choices of the slide are contrasted in a way that is friendly to those with color vision deficiencies.
You can check this using Color Oracle.