Mattingly "AI & Prompt Design: The Basics of Prompt Design"
Recommender system introduction
1. No So Brief Introduction
Case Studies
.
A Study of Recommender System
.
12.05.2011
. . . . . .
A Study of Recommender System
2. No So Brief Introduction
Case Studies
. Main Contents
1. No So Brief Introduction
The categories of RS
The Challenge
Human Recommendation VS. Recommendation System
2. Case Studies
Amazon
YouTube
. . . . . .
A Study of Recommender System
3. No So Brief Introduction
Case Studies
Inforamtion Retrieval ( IR )
google, baidu
Information Filtering ( IF )
strings, facebook
. . . . . .
A Study of Recommender System
4. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
From Web Portal to Search Engine and Personal
. recommendation
. .
Deveplopment Trendency Web portal
. .
Tranditional website
like
. Yahoo! sina
.
Search Engine
.
sometimes hard to
find what you want
.
.
Recommendation
.
“Follow”model
. “cold start”
.
. . . . . .
A Study of Recommender System
5. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Yahoo!
. . . . . .
A Study of Recommender System
6. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Google
. . . . . .
A Study of Recommender System
7. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. YouTube
. . . . . .
A Study of Recommender System
8. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Twitter
. . . . . .
A Study of Recommender System
9. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. What is Recommender System
.
Definition of Recommender System and More examples
.
Recommender System form a specific type of information filtering
system technique that attempts to recommend information items
(movies, TV show, music, news, web pages, research papers etc.)
or social elements (e.g. people, events or groups) that are likely to
be of interest to the user.
.
.
Famous RS website
.
news.qq.com
www.YouTube.com
douban.fm
www.netflix.com
www.taobao.com
.
. . . . . .
A Study of Recommender System
10. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. The Main Job of Recommender System
Top-N problem
. . . . . .
A Study of Recommender System
11. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. An example : Amazon.cn
. . . . . .
A Study of Recommender System
12. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Research Area
1. Recommendation technique and recommendation algorithm
.
2 Real-time research ( users and items are changing )
3. Recommendation quality
.
4 Hybrid data and hybrid technique integration
.
5 Data Mining in RS ( association rule DM , Bayesian
category )
6. User privacy protection
. . . . . .
A Study of Recommender System
13. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Main Contents
1. No So Brief Introduction
The categories of RS
The Challenge
Human Recommendation VS. Recommendation System
2. Case Studies
Amazon
YouTube
. . . . . .
A Study of Recommender System
14. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Recommendation Algorithm
.
Main RS Algorithm
.
.
1 Content-based recommendation
.
2 Collaborative filtering recommendation
user-based CF recommendation ( user CF )
item-based CF recommendation ( item CF )
3. Knowledge-based recommendation
.
4 association-based recommendation
. hybrid recommendation
5
.
. . . . . .
A Study of Recommender System
15. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Association Rule-Based Recommendation ( ARBR )
.
Target: item6
.
f
.rom 1-item set to k-items set
. item5
Procedure
.
.
1 For new user u set a item set
Userb item4
.
2 Find association-rule R for u
u Usera
3. All items right of Ru ⇒ item set item3
.
4 Delete items have been bought
. item2
.
Disadvantage item1
.
H
. ard to find association rules
.
. . . . . .
A Study of Recommender System
16. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Association Rule-Based Recommendation ( ARBR )
.
Target: item6
.
f
.rom 1-item set to k-items set
. item5
Procedure
.
.
1 For new user u set a item set
Userb item4
.
2 Find association-rule R for u
u Usera
3. All items right of Ru ⇒ item set item3
.
4 Delete items have been bought
. item2
.
Disadvantage item1
.
H
. ard to find association rules
.
. . . . . .
A Study of Recommender System
17. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Content-Based Recommendation ( CBR )
.
1. Extract item i ’s attribute features
.
Content(i) = (Wi1 , Wi2 , · · · · · · , Wik )
.
.
2. Compute user c ’s bias
.
ContenBaseProfile( c ) = (Wc1 , Wc2 , · · · · · · , Wck )
Wcj means the important of key word kj to user c.
.
.
3. Matching Key words, filtering low relative items
.
U( c, i ) = score( ContenBaseProfile( c ) , Content( i ) )
hard to extract features from videos and sound.
.
. . . . . .
A Study of Recommender System
18. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Collaborative Filtering Recommendation ( CFR )
.
process
.
} {
U = u1 , u2 , · · · , um predict score
Get score matrix → CF →
I = i1 , i2 , · · · , in top-N R
.
.
flowsheet
.
.
. . . . . .
A Study of Recommender System
19. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Memory-basedUser-based Collaborative Filtering
.
The Process of User-CF
.
Scan all users, find similar interest users, use their scores to predict
the scores of user u.
.
1 Establish a m × n matrix
2. Search neighbors using similarity algorithm
.
3 Recommendation generation
.
. . . . . .
A Study of Recommender System
20. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. similarity algorithm
.
Cosine Function
.
∑n
⃗ ·⃗
a b Rai Rbi
Sim(Ua , Ub ) = cos(⃗ , ⃗ =
a b) = √∑ i=1 √∑
||⃗ || · ||⃗
a b|| n 2 n 2
. i=1 Rai i=1 Rai
.
Pearson Function
.
∑
(RUa ,s − mUa )(RUb ,s − mUb )
s∈Sa,b
Sim(Ua , Ub ) = √ ∑ ∑
(RUa ,s − mUa )2 (RUb ,s − mUb )2
s∈Sa,b s∈Sa,b
.
. . . . . .
A Study of Recommender System
21. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Item-based Collaborative Filtering
.
The Process of Item-CF
.
.
1 Compute similarity among items by using similarity algorithm
.
2 Find several most similar items neighbors of target item
. By weighting items’ scores as the target item’s predicted value
3
.
.
Item prediction by similar items’ scores
.
.
. . . . . .
A Study of Recommender System
22. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. How we predict the Pu,i (user u’s score for item)
∑
∑ j∈NBS sim(i,j)·Ru,j
Pu,i =
j∈NBS (|sim(i,j)|)
NBS: neighbors set of item i, Ru,j : score of user for item j
. . . . . .
A Study of Recommender System
23. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. User-based CF and Item-based CF
.
User-based CF .
. Item-based CF
.
1 User-item matrix .
.
1 User-item matrix
.
2 Find similar users
.
2 Find similar items
3. Compute the similarity of
3. Compute the similarity of
two users’ items
two items for all users
.
4 Recommend similar users’
4. Recommend items which
choices
have high-similarity
5. Asking friends for a
5. People who buy x also buy y
recommendation
.
6 People like similar stuff
.
6 User like items of user who
which they like before
share similar interests .
.
. . . . . .
A Study of Recommender System
24. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. User-based CF and Item-based CF
.
User-based CF .
. Item-based CF
.
1 User-item matrix .
.
1 User-item matrix
.
2 Find similar users
.
2 Find similar items
3. Compute the similarity of
3. Compute the similarity of
two users’ items
two items for all users
.
4 Recommend similar users’
4. Recommend items which
choices
have high-similarity
5. Asking friends for a
5. People who buy x also buy y
recommendation
.
6 People like similar stuff
.
6 User like items of user who
which they like before
share similar interests .
.
. . . . . .
A Study of Recommender System
25. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. User-based CF and Item-based CF
.
User-based CF .
. Item-based CF
.
1 User-item matrix .
.
1 User-item matrix
.
2 Find similar users
.
2 Find similar items
3. Compute the similarity of
3. Compute the similarity of
two users’ items
two items for all users
.
4 Recommend similar users’
4. Recommend items which
choices
have high-similarity
5. Asking friends for a
5. People who buy x also buy y
recommendation
.
6 People like similar stuff
.
6 User like items of user who
which they like before
share similar interests .
.
. . . . . .
A Study of Recommender System
26. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. User-based CF and Item-based CF
.
User-based CF .
. Item-based CF
.
1 User-item matrix .
.
1 User-item matrix
.
2 Find similar users
.
2 Find similar items
3. Compute the similarity of
3. Compute the similarity of
two users’ items
two items for all users
.
4 Recommend similar users’
4. Recommend items which
choices
have high-similarity
5. Asking friends for a
5. People who buy x also buy y
recommendation
.
6 People like similar stuff
.
6 User like items of user who
which they like before
share similar interests .
.
. . . . . .
A Study of Recommender System
27. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. User-based CF and Item-based CF
.
User-based CF .
. Item-based CF
.
1 User-item matrix .
.
1 User-item matrix
.
2 Find similar users
.
2 Find similar items
3. Compute the similarity of
3. Compute the similarity of
two users’ items
two items for all users
.
4 Recommend similar users’
4. Recommend items which
choices
have high-similarity
5. Asking friends for a
5. People who buy x also buy y
recommendation
.
6 People like similar stuff
.
6 User like items of user who
which they like before
share similar interests .
.
. . . . . .
A Study of Recommender System
28. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. User-based CF and Item-based CF
.
User-based CF .
. Item-based CF
.
1 User-item matrix .
.
1 User-item matrix
.
2 Find similar users
.
2 Find similar items
3. Compute the similarity of
3. Compute the similarity of
two users’ items
two items for all users
.
4 Recommend similar users’
4. Recommend items which
choices
have high-similarity
5. Asking friends for a
5. People who buy x also buy y
recommendation
.
6 People like similar stuff
.
6 User like items of user who
which they like before
share similar interests .
.
. . . . . .
A Study of Recommender System
29. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Model-based Collaborative Filtering and other RS
.
Model-based CF RS
More users and items, sparser matrix and worse
recommendation quality
Model from training users’ history, when new user comes, use
this model to predict and recommend
.
. . . . . .
A Study of Recommender System
30. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Model-based Collaborative Filtering and other RS
.
Model-based CF RS
More users and items, sparser matrix and worse
recommendation quality
Model from training users’ history, when new user comes, use
this model to predict and recommend
Demographic-based Recommendation 基于用户统计信息
Categorize users based on personal attributes ( ASL,... ) and
make recommendation based on demographic classes
.
. . . . . .
A Study of Recommender System
31. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Model-based Collaborative Filtering and other RS
.
Model-based CF RS
More users and items, sparser matrix and worse
recommendation quality
Model from training users’ history, when new user comes, use
this model to predict and recommend
Demographic-based Recommendation 基于用户统计信息
Categorize users based on personal attributes ( ASL,... ) and
make recommendation based on demographic classes
Utility-based Recommendation 基于效用
.
. . . . . .
A Study of Recommender System
32. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Model-based Collaborative Filtering and other RS
.
Model-based CF RS
More users and items, sparser matrix and worse
recommendation quality
Model from training users’ history, when new user comes, use
this model to predict and recommend
Demographic-based Recommendation 基于用户统计信息
Categorize users based on personal attributes ( ASL,... ) and
make recommendation based on demographic classes
Utility-based Recommendation 基于效用
Knowledge-based Recommendation
Inference technique 推理技术
.
. . . . . .
A Study of Recommender System
33. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Model-based Collaborative Filtering and other RS
.
Model-based CF RS
More users and items, sparser matrix and worse
recommendation quality
Model from training users’ history, when new user comes, use
this model to predict and recommend
Demographic-based Recommendation 基于用户统计信息
Categorize users based on personal attributes ( ASL,... ) and
make recommendation based on demographic classes
Utility-based Recommendation 基于效用
Knowledge-based Recommendation
Inference technique 推理技术
Hybrid Recommendation
.
. . . . . .
A Study of Recommender System
34. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. A Example of Demographic-based Recommendation
.
Refrence: 动态推荐系统关键技术研究.xlvector
.
Male users is more than female users in IMDB
No Cold Start
Problem: the recommendation is rough
.
. . . . . .
A Study of Recommender System
35. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. The Comparison of Main Recommendation System
.
RS Advantage Disadvantage
关联规则推荐 发现新兴趣点,不 关联规则难以获
需要领域知识 取,个性化程度低
基于内容推荐 推荐结果简单,不 新用户问题,需要
需要领域知识 足够对象构造分类
器
协同过滤推荐 跨类兴趣推荐,自 “cold start”,稀疏
动化程度高,处理 性问题,需大量历
非机构化对象 史数据
基于用户信息 发现新兴趣点,不 获得人口统计信息
需领域知识 难度大
基于知识推荐 把用户的需求映射 知识难以获得,推
到产品上,考虑非 荐是静态的
. 产品属性
. . . . . .
A Study of Recommender System
36. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Main Contents
1. No So Brief Introduction
The categories of RS
The Challenge
Human Recommendation VS. Recommendation System
2. Case Studies
Amazon
YouTube
. . . . . .
A Study of Recommender System
37. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. 5 Problems of Recommender Systems (1st)
.
Lack of Data
.
firstly needs item data,then capture and analyze user data
excellent recommendations are those with lots of consumer
user data. Such as Amazon Google
.
.
Changing Data
.
too many product attributes in fashion and each attribute has
a different level of importance at different times for the same
consumer
the trends are always changing
.
. . . . . .
A Study of Recommender System
38. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. 5 Problems of Recommender Systems (2nd)
.
Changing User Preference
.
books for myself or a birthday present for my sister
.
.
Unpredictable Items
.
the user reaction to items tends to be diverse and
unpredictable
.
.
This Stuff is Complex
.
takes a lot of variables to do even the simplest
recommendations
Netflix looking for a 10% improvement on their algorithm
.
. . . . . .
A Study of Recommender System
39. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Recommendation In Reality ( 1st )
.
Real Recommendation System must satisfy following demands
.
RS must fullfill users’ demands
The data generated is beneficial for RS’s development
RS must fullfill owner’s demands
.
. Data Recommender Data
Customer Owner
Rec. system Data
. . . . . .
A Study of Recommender System
40. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Recommendation In Reality ( 2nd )
.
What users want ?
.
Good Recommendation from Recommender System
.
.
How do we judge a RS ?
.
accuracy — recommend items that users like
coverage — long tail effect (personalization)
diversity — recommend kinds of items
novelty and serendipity — user experience is important
.
.
What owner will consider ?
.
long tail effect
sale specific goods — (Promotion,unsaleable goods)
.
. . . . . .
A Study of Recommender System
41. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Recommendation In Reality ( 3rd )
.
Society bias and user bias shifting
.
.
. . . . . .
A Study of Recommender System
42. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Recommendation In Reality ( 4th )
.
Item bias shifting
.
.
. . . . . .
A Study of Recommender System
43. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Recommendation In Reality ( 5th )
.
Cold Start is a problem, But we can:
.
Hybrid RS combining
association-basedcontent-baseddemographic-based RS
Social Recommendation (Ask friends for recommendations)
Amazon makes Social Recommendation using Facebook
Connector ( API )
.
.
What we know more ?
.
If all users rely on a Recommender System, the recommendation
algorithm’s accuracy will be 100%
Even recommendation algorithm doesn’t improve, the accuracy will
rise.
.
. . . . . .
A Study of Recommender System
44. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Main Contents
1. No So Brief Introduction
The categories of RS
The Challenge
Human Recommendation VS. Recommendation System
2. Case Studies
Amazon
YouTube
. . . . . .
A Study of Recommender System
45. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Reference
Comparing Recommendations Made by Online Systems
and Friends
. . . . . .
A Study of Recommender System
46. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Paper’s Study Goal,Study design,Study method
.
What we know after reading this paper
.
Goal:paper compares RS by online RS and friends.
book RS: Amazon, RatingZone, Sleeper
movie RS: Amazon, MovieCritic, Reel.com
.
. . . . . .
A Study of Recommender System
47. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Paper’s Study Goal,Study design,Study method
.
What we know after reading this paper
.
Goal:paper compares RS by online RS and friends.
book RS: Amazon, RatingZone, Sleeper
movie RS: Amazon, MovieCritic, Reel.com
Definition:
good R: interest the user
useful R: user’s interest, no experienced before
trust-generating R: positive experiences previously
.
. . . . . .
A Study of Recommender System
48. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Paper’s Study Goal,Study design,Study method
.
What we know after reading this paper
.
Goal:paper compares RS by online RS and friends.
book RS: Amazon, RatingZone, Sleeper
movie RS: Amazon, MovieCritic, Reel.com
Definition:
good R: interest the user
useful R: user’s interest, no experienced before
trust-generating R: positive experiences previously
Method:
participants:19 people,20 to 35 years,6 males 13 females
procedure:complete registration; rated items on each RS in
order to get recommendations; review list of recommendations;
complete satisfaction and usability questionaire. three frends
recommend 3 books and movies which the user haven’t
discussed with friends before.
. . . . . . .
A Study of Recommender System
49. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Conclusion
.
1. User prefer recommendation made by their friends to those
made by the set of online RS.
.
2 User expressed a high level of overall satisfaction with online
RS.
3. Design recommendations for RS:
user don’t mind rating more items initially to receive quality
recommendation
Allow users to provide initial rating on a continuous rather
than binary choice scale
provide enough information about the recommended item for
user to make a decision
provide easy ways to generate new recommendation sets
Interface matters, mostly when it gets in the way
.
. . . . . .
A Study of Recommender System
50. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Conclusion
.
1. User prefer recommendation made by their friends to those
made by the set of online RS.
.
2 User expressed a high level of overall satisfaction with online
RS.
3. Design recommendations for RS:
user don’t mind rating more items initially to receive quality
recommendation
Allow users to provide initial rating on a continuous rather
than binary choice scale
provide enough information about the recommended item for
user to make a decision
provide easy ways to generate new recommendation sets
Interface matters, mostly when it gets in the way
.
. . . . . .
A Study of Recommender System
51. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Conclusion
.
1. User prefer recommendation made by their friends to those
made by the set of online RS.
.
2 User expressed a high level of overall satisfaction with online
RS.
3. Design recommendations for RS:
user don’t mind rating more items initially to receive quality
recommendation
Allow users to provide initial rating on a continuous rather
than binary choice scale
provide enough information about the recommended item for
user to make a decision
provide easy ways to generate new recommendation sets
Interface matters, mostly when it gets in the way
.
. . . . . .
A Study of Recommender System
52. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Conclusion
.
1. User prefer recommendation made by their friends to those
made by the set of online RS.
.
2 User expressed a high level of overall satisfaction with online
RS.
3. Design recommendations for RS:
user don’t mind rating more items initially to receive quality
recommendation
Allow users to provide initial rating on a continuous rather
than binary choice scale
provide enough information about the recommended item for
user to make a decision
provide easy ways to generate new recommendation sets
Interface matters, mostly when it gets in the way
.
. . . . . .
A Study of Recommender System
53. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Conclusion
.
1. User prefer recommendation made by their friends to those
made by the set of online RS.
.
2 User expressed a high level of overall satisfaction with online
RS.
3. Design recommendations for RS:
user don’t mind rating more items initially to receive quality
recommendation
Allow users to provide initial rating on a continuous rather
than binary choice scale
provide enough information about the recommended item for
user to make a decision
provide easy ways to generate new recommendation sets
Interface matters, mostly when it gets in the way
.
. . . . . .
A Study of Recommender System
54. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Conclusion
.
1. User prefer recommendation made by their friends to those
made by the set of online RS.
.
2 User expressed a high level of overall satisfaction with online
RS.
3. Design recommendations for RS:
user don’t mind rating more items initially to receive quality
recommendation
Allow users to provide initial rating on a continuous rather
than binary choice scale
provide enough information about the recommended item for
user to make a decision
provide easy ways to generate new recommendation sets
Interface matters, mostly when it gets in the way
.
. . . . . .
A Study of Recommender System
55. The categories of RS
No So Brief Introduction
The Challenge
Case Studies
Human Recommendation VS. Recommendation System
. Conclusion
.
1. User prefer recommendation made by their friends to those
made by the set of online RS.
.
2 User expressed a high level of overall satisfaction with online
RS.
3. Design recommendations for RS:
user don’t mind rating more items initially to receive quality
recommendation
Allow users to provide initial rating on a continuous rather
than binary choice scale
provide enough information about the recommended item for
user to make a decision
provide easy ways to generate new recommendation sets
Interface matters, mostly when it gets in the way
.
. . . . . .
A Study of Recommender System
56. No So Brief Introduction Amazon
Case Studies YouTube
. Main Contents
1. No So Brief Introduction
The categories of RS
The Challenge
Human Recommendation VS. Recommendation System
2. Case Studies
Amazon
YouTube
. . . . . .
A Study of Recommender System
57. No So Brief Introduction Amazon
Case Studies YouTube
. Reference
Amazon.com Recommendations Item-to-Item Collaborative
Filtering
. . . . . .
A Study of Recommender System
58. No So Brief Introduction Amazon
Case Studies YouTube
. Amazon’s Demand and Solution
.
Demand
.
Amazon has more than 29 million customers and several
million catalog items
Amazon use recommendation algorithms to personalize the
online store for each customer in real time.
.
.
Solution
.
Existing algorithms were evaluated over small data sets.
. Reduce M by randomly sampling the customers or discarding
1
customers with few purchases (M:the number of customers)
.
2 Reduce N by discarding very popular or unpopular items (N:the
number of items)
.
3 Dimensionality reduction techniques such as clustering and
principal component analysis can reduce M or N
.
. . . . . .
A Study of Recommender System
59. No So Brief Introduction Amazon
Case Studies YouTube
The Amazon Item-to-Item Collaborative Filtering
. Algorithm
.
For each item in product catalog, I1
For each customer C who purchased I1
For each item I2 purchased by customer C
Record that a customer purchased I1 and I2
For each item I2
. Compute the similarity between I1 and I2
.
Algorithm Complexity
.
Worst case: O(N2 M)
In practice: O(NM) , ’cause customers have few purchases
Sampling customers who purchase best-selling titles reduces
runtime even further
.
. . . . . .
A Study of Recommender System
60. No So Brief Introduction Amazon
Case Studies YouTube
. Main Contents
1. No So Brief Introduction
The categories of RS
The Challenge
Human Recommendation VS. Recommendation System
2. Case Studies
Amazon
YouTube
. . . . . .
A Study of Recommender System
61. No So Brief Introduction Amazon
Case Studies YouTube
. From ten pages to four pages
.
Reference papers
.
Shumeet Baluja etc.,2008,Video Suggestion and Discovery for
YouTube: Taking Random Walks Through the View Graph.
James Davidson etc.,2010,The YouTube Recommendation
System.
.
. . . . . .
A Study of Recommender System
62. No So Brief Introduction Amazon
Case Studies YouTube
. Taking Random Walks Through the View Graph (1st)
.
Random Walk in multi-demension
.
some walks take their steps at random times
a fundamental topic in discussions of Markov processes
.
. . . . . .
A Study of Recommender System
63. No So Brief Introduction Amazon
Case Studies YouTube
. Taking Random Walks Through the View Graph (2nd)
.
Co-view Graph
.
Each video is a vertex
in the graph that is
linked to other videos
.
.
The Adsorption Algorithm
.
classify a node in a graph in terms of labels present on some
of the other nodes.
. u and v have a short path between them
1
.
2 u and v have several paths between them
.
3 u and v have paths that avoid high-degree nodes
.
. . . . . .
A Study of Recommender System
64. No So Brief Introduction Amazon
Case Studies YouTube
. Taking Random Walks Through the View Graph (3rd)
.
Adsorption via Averaging
.
Input: G = (V, E, w), L, VL
repeat
∪˜
for each v ∈ V∑ V do:
Let Lv = u w(u, v)Lu
end-for
Normalize Lv to have unit L1 norm
until covergence
Output: Distributions Lv |v ∈ V
.
. . . . . .
A Study of Recommender System
65. No So Brief Introduction Amazon
Case Studies YouTube
. Taking Random Walks Through the View Graph (4th)
.
Adsorption via Random Walks
.
Input: G = ∪ E, w),L,VL , distinguished vertex v:
(V,
˜ ˜ ∪ v
Let G = (V V, E (v, ˜)|v ∈ VL , w).
Define w(v, ˜) = 1 for all v ∈ VL
v
done := false
vertex := v
while ( not done ) do:
vertex := pick-neighbor(v, E, w)
˜
if(neighbor ∈ V)
done := true
end-while
u := vertex
Output: label according to Lu .
.
. . . . . .
A Study of Recommender System
66. No So Brief Introduction Amazon
Case Studies YouTube
. Taking Random Walks Through the View Graph (5th)
.
Adsorption via Linear Systems
.
Input: G= (V, E, w)
Let n:=|V|
Define the linear system of equations in n2 variables Xuv ,
for u, v ∈ V: ∑
v Xuv = 1 ∀u ∈ V;
∑
z:(z,u)∈E w(z, u)Xuv = Xuv ∀u, v ∈ V.
incremental update to the label distributions or addition or
deletion of nodes can be easily accommodated by quickly
updating the information for the relevant neighborhood of the
graph.
.
. . . . . .
A Study of Recommender System
67. No So Brief Introduction Amazon
Case Studies YouTube
. The YouTube Video Recommendation System (1st)
.
Challenges
.
No or very poor metadata
User interactions are relatively short and noisy
.
.
System Design
.
recommend recent and fresh as well as diverse and relevant to
user’s actions.
generated by using a user’s personal activity.
.
.
Input Data
.
Content data
User activity data (explicit data , implicit data)
.
. . . . . .
A Study of Recommender System
68. No So Brief Introduction Amazon
Case Studies YouTube
. The YouTube Video Recommendation System (2nd)
.
Related Videos
.
.
1 A given time period, we count for each pair of videos v ,v how
i j
often they were co-watched.
.
2 We denote this co-visitation count by c j.
i
3. We define related score of video vj to vi as :
c
r(vi , vj ) = f(vi ij j )
,v
.
4 f(v , v ) is a normalization function that denote the global
i j
popularity
.
5 Pick the set of related videos R for a given seed video v as
i i
the top N candidate videos ranked by their scores r(vi , vj ).
.
. . . . . .
A Study of Recommender System
69. No So Brief Introduction Amazon
Case Studies YouTube
. The YouTube Video Recommendation System (3rd)
.
Generating Recommendation Candidates
.
.
1 A given seed set S (e.g. the videos user watched)
.
2 Each video v in the seed set consider its related videos R
i i
.
3 Denote the union of these related video sets as C :
1
∪
C1 (S) = Ri
vi ∈CS
. Distance of n from any video in the seed set:
4
∪
Cn (S) = Ri
vi ∈Cn−1
.
. . . . . .
A Study of Recommender System
70. No So Brief Introduction Amazon
Case Studies YouTube
. The YouTube Video Recommendation System (4th)
.
Generating Recommendation Candidates
.
The final candidate set Cfinal :
∪
N
Cfinal = ( Ci ) − S
i=0
Due to the high branching factor of the related videos graph
we found expanding over a small distance yielded broad and
diverse recommendations even for users with a small seed set.
.
.
Ranking
.
Using a linear combination of three kinds of signals( a.video
quality b.user specifility c.diversification ),we generate a ranked
list
. of the candidate videos.
. . . . . .
A Study of Recommender System