Transcript ICISO 2015

An Optimization of Collaborative Filtering Personalized
Recommendation Algorithm Based on Time Context Information
Xian Jin, Qin Zheng and Lily Sun
ICISO 2015, Toulouse
2015-03-20
Au t h o r s
Xian JIn
PhD. Candidate of Shanghai University of Finance and Ecnomics
Major in Management Science and Engineering
MBA and Software Engineering
Worked in Tencent（2 years）、Autodesk（6 years）.
Prof. Qing Zheng
vice-president of South University of Science and Technology of China (SUSTC)
CPC member, a doctorate degree holder, professor, and doctoral supervisor.
Dr. LiLy Sun
PhD. Candidate of Shanghai University of Finance and Ecnomics
Information and Communications Technology Research
The university of Reading
1
ICISO 2015
1
Why we need personalized recommendation？
Where the application of personalized recommendation？
2
CONTENTS
3
4
5
2
How is the personalized recommendation works？
What I did in this paper?
About the Future
ICISO 2015
[1]
3
Why
we need personalized recommendation？
ICISO 2015
4
ICISO 2015
5
ICISO 2015
With the explosive growth of the Internet information, users facing
serious problems of information overload in big data times. The user
needs to spend a lot of time and effort to find useful information.
Base on this background, personalized recommendation system emerge
as the times require, and it can help users to acquire useful information
and knowledge from the massive information.
That’s why the personalized recommendation appears in our life.
6
ICISO 2015
[2]
7
Where
the application of personalized recommendation？
ICISO 2015
Personalized recommendation can be find in many e-commerce website.
Amazon
Inspired by your shopping trends
Inspired by your browsing history
Customers with similar searches purchased
Viewed this also viewed
Bought this also bought
Ultimately buy after viewed this
Customers who bought items in your cart
also bought
Products with similar tags
Today’s recommendations for you
New for you
8
ICISO 2015
And some movie discover website.
SynopsiTV:
Personalized movie recommendation platform
9
ICISO 2015
And some daily living website.
Hunch local(LB)
• Search results
• Like
• Dislike
• Unique V.S popular
• Question based
10
ICISO 2015
There are lots of models and algorithms
in the recommendation system.
11
ICISO 2015
[3]
12
How
is the personalized recommendation works？
11.11
Single's Day
13
ICISO 2015
After 2009-11-11, this day was re-defined
By Alibaba Inc.
11.11
Online Shopping Day = Black Friday online
ICISO 2015
Single's Day
14
11.11
What is the 57.112 billion means？
0.08% of Chinese GDP of Y2014
15
ICISO 2015
ICISO 2015
How many sales of 11.11 comes from personal recommendation?
Personal recommendation Sales
in alibaba 11.11 (Billions)
30
25.19
2014 Alibaba 11.11
Recommendation Sales percentage of
total sales:
30%
25
Recommendation caused User number:
20
92.61 Millon
15
10
8.43
5
0.5
0
2012
16
2013
2014
ICISO 2015
Here is the accuracy difference between four algorithms
Global ranking (Blue),
Collaborative filtering (yellow),
Heat conduction (purple)
Diffusion (red)
17
ICISO 2015
In recommendation area, there mainly are ten hot research topics.
数据稀疏性问题
 The problem of sparse data
冷启动问题
 The problem of cold start
大数据处理与增量计算问题
 The big data processing and incremental calculation problem
多样性与精确性的两难困境
 The dilemma of the diversity and accuracy
推荐系统的脆弱性问题
 The vulnerability of recommendation systems
用户行为模式的挖掘和利用
 Mining and utilization of user behavior mode
推荐系统效果评估
 Recommendation system effectiveness evaluation
用户界面与用户体验
 The user interface and user experience
多维数据的交叉利用
 Cross utilization of multidimensional data
社会推荐
 Social recommendation
18
ICISO 2015
I want do more research here，and try to improve the diversity and accuracy.
The dilemma of the diversity and accuracy
You know, recommend system was calculated based on the history data. So the recommend list must be the items which
user was like, most of research doing well in accruracy.
But human always like the fresh items, maybe something they never click\search\buy.
So a good recommend model should balance the diversity and accuracy.
Something like “Create a fresh accuracy requirement for Users”
19
ICISO 2015
[4]
20
What
I did in this paper?
ICISO 2015
In this paper I made some optimize on CF method.
What is the Collaborative filtering approach?
21
Base on User
Base on Item
ICISO 2015
CF method do have mainly four shortages.
（1）The problem of sparse data
（ 2）Method hardly to extend。
（ 3）The problem of cold start ，such as new users and new items。
（ 4）hardly to generate Fresh interest-------Scenarios
22
ICISO 2015
Add time division to create a Users-Items-Time Matrix.
Time was divided as Monday ….. Sunday; Mornig-noon-evening…
User will get different recommend lists in different time.
This method can improve both the diversity and accuracy
23
ICISO 2015
The result is better than normal method.
24
ICISO 2015
[5]
25
About the future
ICISO 2015
Time division is just a start…
Users….Items….Time…Location….Weather….Doing….Social connect….
The future is coming….
26