This document discusses recommendation services used by Amazon and how they could be implemented on 123Mua.vn. It describes Amazon's business model for recommendations, including recommending products based on browsing history, viewing history, and purchases. It also discusses the research model, including content-based filtering, collaborative filtering, and how they calculate similarity. Finally, it proposes how 123Mua.vn could implement a recommendation engine using items of known interest, popular items, similar item lists, and a weighting scheme to generate recommendations.
Recommendation Systems: Applying Amazon's Collaborative Filtering Methods to E-commerce
1. Amazon Recommendation Services 123Mua.vn Recommendation Services
Recommendation Services
Nguyen.Cao-Duc
Data Mining Team Lead
E-Commerce & Services Dept.
VNG Corp.
Internal Research Document
September 19, 2012
2. Amazon Recommendation Services 123Mua.vn Recommendation Services
Outline
1 Amazon Recommendation Services
Business Model
Research Model
Implementation Model
2 123Mua.vn Recommendation Services
Business Model
Implementation Model
4. Amazon Recommendation Services 123Mua.vn Recommendation Services
Business Model
Recommendation Services (cont.)
Browsing Product Recommendations:
5. Amazon Recommendation Services 123Mua.vn Recommendation Services
Business Model
Recommendation Services (cont.)
Viewing Product Recommendations:
6. Amazon Recommendation Services 123Mua.vn Recommendation Services
Business Model
Recommendation Services (cont.)
Purchasing Product Recommendations:
7. Amazon Recommendation Services 123Mua.vn Recommendation Services
Business Model
Recommendation Services (cont.)
How to have such Recommendations:
8. Amazon Recommendation Services 123Mua.vn Recommendation Services
Business Model
Recommendation Services (cont.)
Other Recommendations:
9. Amazon Recommendation Services 123Mua.vn Recommendation Services
Research Model
Recommendation Problem
The main purpose of the recommendation system is to
recommend personalized products to users of a merchant’s
Web site.
Two types of Recommendations:
Content-based Filtering
Recommend items with similar content.
Collaborative Filtering
Recommend items based on interests of a community of
users.
Hybrid Content-based Collaborative Filtering
Combination the two above approaches to overcome the
disadvantages of each approach.
11. Amazon Recommendation Services 123Mua.vn Recommendation Services
Research Model
Collaborative Filtering - Problem Description
Question: What should be the rating of Sam for Yellow?
Approach: Use ratings of other users (user-based CF) or
other items (item-based CF)
13. Amazon Recommendation Services 123Mua.vn Recommendation Services
Research Model
Similarity Computation
Vector Cosine-based Similarity:
Formular:
wu,v = i∈I ru,irv,i
i∈I r2
u,i i∈I r2
v,i
wi,j = u∈U ru,iru,j
u∈U r2
u,i u∈U r2
u,j
where:
I is the set of items that both user u and v have rated.
U is the set of users who rate both item i and i.
Drawbacks
Different users have their own rating scales.
14. Amazon Recommendation Services 123Mua.vn Recommendation Services
Research Model
Similarity Computation (cont.)
Correlation-based Similarity:
Formular:
wu,v = i∈I(ru,i − ¯ru)(rv,i − ¯rv )
i∈I(ru,i − ¯ru)2
i∈I(rv,i − ¯rv )2
wi,j = u∈U(ru,i − ¯ri)(ru,j − ¯rj)
u∈U(ru,i − ¯ri)2
u∈U(ru,j − ¯rj)2
where:
I is the set of items that both user u and v have rated.
U is the set of users who rate both item i and i.
15. Amazon Recommendation Services 123Mua.vn Recommendation Services
Research Model
Collaborative Recommendation
Given a user u:
User-based prediction:
Aggregate the ratings of other users:
Pu,i = ¯ru + v∈V (rv,i − ¯rv )wu,v
v∈V |wu,v |
where V is the set of all users have rated the item i
Item-based prediction:
Simple weighted average:
Pu,i = n∈N ru,nwi,n
n∈N |wi,n|
where N is the set of other rated items of user u
16. Amazon Recommendation Services 123Mua.vn Recommendation Services
Research Model
Collaborative Filtering - Drawbacks
User has to rate items to build profiles as well as item has
to be rated (cold-start problem: new user, new item, new
system)
Recommendations may not be diversed
18. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Recommendation Service Components
Recommendation Service Components takes Items of
Known Interest of the given User and Similar Items Table to
create Recommendation Items.
20. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Recommendation Service Components (cont.)
Sources of Items of Known Interest with respect to a User:
User shopping card activities
User purchasing activities
User favorite items profile (i.e WishList)
Popular items are items satisfied some pre-specified popular
criteria:
Number of item views
Time on item view
Number of item purchasings
21. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Recommendation Service Components (cont.)
Each cell in the Similar Items Table associates a commonality
index (CI) to indicate the relatedness of that item with the
popular item. The relatedness of two items i, j could be
expressed via:
Two items have been purchased together
Two items have been rated similarly
or the value of wi,j = u∈U (ru,i −¯ri )(ru,j −¯rj )
√
u∈U (ru,i −¯ri )2
√
u∈U (ru,j −¯rj )2
or the similarity between two items using content-based
filtering
or . . . combinations of all above with some controlled
parameters.
22. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Recommendation Service Components (cont.)
Similar lists are combined appropriately by a weighting
scheme representing the relative importance of popular
items with respect to the items of known interests.
Weighting scheme of similar item lists:
Rating of the user to the popular item.
User purchased multiple copies of the popular item
Time user spend on the popular item
Recent purchasing items are weighted more than earlier
purchasing items
23. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Recommendation Service Components (cont.)
The combined sorted list of similar items lists may need be
modified to remove certain items:
Items already purchased or rated by user or have been
viewed by user and its content has not changed.
Items not in any product groups registered by user.
The combined sorted list of similar items lists may need be
modified to add certain items:
Items user has considered to purchase but did not
purchase.
Items user has viewed but its content has changed after
that.
The recommendation result may be transfered to the end
user by different types of transmission methods (view on
site, email, mobile message, chat message, etc.)
31. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Items of Known Interest
Sources of Items of Known Interest with respect to a User:
User category browsing or item viewing activities
User shopping card activities
33. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Popular Items (cont.)
Popular items are items satisfied some pre-specified popular
criteria:
Number of page views on an item and/or category of the
item and/or shop of item
Time on an item and/or category of item and/or shop of
item
Bounce Rate on that item
Exit Rate on that item
36. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Similar Item List (cont.)
Sources of Similar Item List by targeting to certain items
based on:
Number of page views on an item and/or category of the
item and/or shop of item
Time on an item and/or category of item and/or shop of
item
Bounce Rate on that item
Exit Rate on that item
or items follow certain business objectives such as items
is Up within a period of time.
Sources of Similar Item List with respect to a Popular item:
Items of the same category and/or shop
37. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Similar Item List (cont.)
Each item in the Similar Item List associates a commonality
index (CI) to indicate the relatedness of that item with the
popular item. The relatedness of two items i, j could be
expressed via:
The value of wi,j = u∈U (ru,i −¯ri )(ru,j −¯rj )
√
u∈U (ru,i −¯ri )2
√
u∈U (ru,j −¯rj )2
where:
ru,i represents interest level of user u towards item i
or the similarity between two items using content-based
filtering
or . . . combinations of all above with some controlled
parameters.
38. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Weighting Scheme of Similar Item Lists
Similar item lists are combined appropriately by a
weighting scheme representing the relative importance of
popular items with respect to the items of known interests.
Weighting scheme of similar item lists of popular items:
User views the popular item multiple times
Time user spent on the popular item
Recent viewings of popular items weighted more
39. Amazon Recommendation Services 123Mua.vn Recommendation Services
Implementation Model
Question
How to identify popular topics from multiple
related/independent properties?
How to measure the interest of a topic viewed by a user?