1. About Our Recommender System
Kimikazu Kato, Chief Scientist
Silver Egg Technology Co., Ltd.
2. Table of Contents
• About myself
• About the company and its business
• Survey on related researches
• Conclusion
1
3. About Myself
Kimikazu Kato
• Ph.D in computer science, background in
mathematics
• Joined Silver Egg as a Chief Scientist in Nov. 2012
• Experiences in numerical computation
– 3D CAD, geometric computation
– Computer graphics
– Partial differential equation
– Parallel computation, GPGPU
• Now designing the core of recommender system
2
4. About Silver Egg Technology
Silver Egg Technology Book written by CEO
Established: September, 1998
CEO: Tom Foley ≪著書≫「One to Oneマーケティングを超えた
COO: Junko Nishimura 戦略的Webパーソナライゼーション」
(出版社:日経BP社 発売:2002年5月)
Capital: ¥78 Million
Main Services: 「ASP・SaaS・ICTアウトソーシングアワード2009」
ASP・SaaS部門「委員長特別賞」受賞
Recommender System
Online Advertisement Service
第8回(2010)、第9回(2011)
「デロイト 日本テクノノロジー Fast50」受賞
3
5. Recommender System
Recommender system proposes the items best fit for individuals’ needs.
Good recommender system provides a comfort for online shopping
experiences and improves customer loyalty.
Ranking
No.1 No.2 No.3
XXXXXX
XXXXXXXXX
3,800円
Combination
Additional
Cross-sell
4
6. Consistent behavior targeting
Consistent user behavior targeting from “traffic inflow” to “retention” is
essential for improving sales and profit.
Pre-access On-access Post-access
Traffic inflow Service Conversion Retention
Recommender Aigent Mail
Aigent
HotView Aigent Aigent Gadget Aigent On-
Personalized Aigent Transaction Recommender Event Driven
Retargeting ad LPO Recommado Portal Demand Printing
Search Recommender Mail Mail Mail
Aigent Suite (Real Time Recommender Platform)
Silver Egg Technology provides smart targeting technology
which enables optimization of online marketing
5
7. Interaction of Advertisement and Recommender
Media Dashboard
Merchandizer
-Registers items to promote
Consumer - Checks performance
Discovery in a
media site
Shows ads of items to
promote to the target
users
HotView - Timestamp
- Geographic information
- Use behavior
- Demands
To the shopping site - Contexts (search words)
百貨店
Aigent
To the site they are interested 通販カタログ Aigent Suite
ブティック
Recommendation for
TVショッピング up-sell and cross-sell
アパレル
Retailer
Ad contents based on users behaviors in shopping sites are more likely to attract
attentions and effectively lead users back to those sites
6
8. Mechanism
Aigent server
Client’s EC site
“Who bought what”
Stored and analyzed
“Who is browsing what”
Respond in real time
“What should be recommended”
ASP service +
Batch update of inventory
Code snippet to connect
with AIgent
Characteristics:
• Real time response
• Implemented as an add-on (cost efficient)
7
9. Consulting Services
• Just showing the result of mathematical
computation is not enough
• To extract optimal sales, parameters should be
tuned by hand
– Statistical co-relation is not all that matters.
• Sometimes recommendations should reflect some
“intention”
– According to policy, strategy, etc.
• Continuous monitoring and A/B testing
8
10. About recommendation algorithms
• Collaborative filtering
• Fruitful methods as a result of Netfilx Prize
– Neighborhood Models
– Matrix factorization
– Restricted Boltzmann Machines
9
11. Netflix Prize
The Netflix Prize was an open competition for the best collaborative filtering
algorithm to predict user ratings for films, based on previous ratings
— Wikipedia
Netflix provided open data for this competition
Closed in 2009
10
12. Movie Rating Prediction
Each user gives rating to the movies they saw
movie
user W X Y Z
A 5 4 1 4
B 4
C 2 3
D 1 4 ?
Is it possible to predict the rating of unknown user/movie pair?
Ratings are expressed as a sparse matrix.
A zero value of the matrix doesn’t really mean “zero” but “unknown”
11
13. Probabilistic Matrix Factorization
Regarding ratings are expressed by small number of components
𝐴 𝑈𝑇 𝑉 noise
Approximate only the non-zero elements
12
15. Rating vs Purchase
Movie rating Purchase recommendation
movie item
user W X Y Z user W X Y Z
A 5 4 1 4 A 1 1 1 1
B 4 B 1
C 2 3 C 1
D 1 4 ? D 1 1 ?
Predicts the rating for the user and Predicts how likely the user buy the
movie pair. item
The matrix includes negative feedback No negative feedback
(Some movies are rated as “boring”) (No reason is given for missing elements)
=> Strong bias toward 1
Only one kind of value for known elements
=> Gives more degree of freedom
A method successful in movie rating prediction is not
useful for recommendation of usual shopping site.
14
16. Solutions
• Regard a zero element as a negative feedback
– Too ad hoc but better than naïve PMF
• Assume a certain ratio of zero elements becomes
one at the optimum [Sindhwani et al. 2010]
– Assign other variables to zero elements and solve a
relaxed optimization
– Experimentally outperform the “zero-as-negative”
method.
V.Sindhwani et al., One-Class Matrix Completion with Low-Density Factorizations. In Proc. of ICDM
2010: 1055-1060
15
18. Conclusion
• Scientific approach is important
– Math really makes money
• But that alone is not enough for real business
• Engineering matters
– Efficient platform and easy-to-deploy mechanism
• Hand tuning part always remains
– Consulting for parameter tune is essential
17