Polkadot JAM Slides - Token2049 - By Dr. Gavin Wood
Atividades de Pesquisa da Rakuten para o E-commerce Global - Satoshi Sekine / Diretor do Instituto de Tecnologia da Rakuten NY
1. 1
Recent Activities for the Rakuten
Global e-commerce
Atividades de Pesquisa da Rakuten
para o e-commerce Global
Rakuten Institute of Technology, New York
2. 2
Rakuten Institute of Technology
Turning emerging and growing new technology seeds
into new business/service opportunities
to enrich the internet life (& real life) all over the world
Mission
The Next Reality
- New reality through emerging technologies -
Concept Tokyo & NY
Strategic R&D organization for Rakuten group
3. 3
Our goal
How was the
carrot you got
yesterday?
Today this fresh
lettuce is good
Recommendation
You will take
a vacation?
The travel agent
over there gives
good service
Affiliation
You are on diet,
but like sweet
food?
This sweet tomato
is good for your
health
Personalization
Old-fashioned vegetable shop
owner
4. 4
Our goal
You want to
know the
vegetable in
this photo…
It is an artichoke.
Image
recognition
You can boil it
and eat it with
mayonnaise
World knowledge
Recently, it is very
popular among
young people
Opinion mining
Old-fashioned vegetable shop
owner
5. 5
Create happy shopping experiences
Hints from a legendary & successful owner at a market
Understand language
Master his products
Understand the customers
World Knowledge
Recognize image, video
Remember for the future
Manipulate knowledge
Think, inference, analysis
Situation adaptation
Friendly service
NLP
Multi-media
Data
Semantic
Data
I/F
AI
Infrastructure
10. 10
Global vs. Local
Problem: Culture/Market dependency
We need one uniform DB structure = Global Catalog
Organize globally Adapt Locally
Sports Equipment
PelotaBaseball
11. 11
Overview
attribute value
Product information by
Attributes and values
Product types
Product category for display is
decided by “product types” and a set
of attribute values
Mapping
R-HC Local Hierarchy
18. 18
Actual Contributions
0
10
20
30
40
Jul. Aug. Sep. Oct. Nov.
GMS
0
10
20
30
40
Jul. Aug. Sep. Oct. Nov.
GMS
(1) 2 big merchants
2 million products
(2011/9)
FIX Merchant A FIX
2011 2011
(3) Expanding to more categories
(2) 2 categories, which include 1,000 merchants
1 million products (2011/11)
Merchant B
→ Contributed to GMS
5 million products (2012/10)
20. 20
3. Automatic Attribute Extraction
Unsupervised Extraction of
Attribute and value from
product description
Attribute Value
Color Red
County Chile
Grape Cabernet franc,
Merlot,
Vintage 2009
22. 22
Synonym Detection
• Attribute names that share a popular value and
don’t appear in the same table are synonym
Region
France, German, Italy …
Alcohol
District
Temperature
for drinking
15 degree, 8 degree …
Synonym
Not Synonym
23. 23
KB for wine category
Grape Type Volume Region Winery Taste
Chardonnay 750ML France Farnese Dry
Chardonnay
100%
720ML Italy Mas de
Monistrol
Full body
Merlot 375ML Spain Leroy Medium body
Riesling 500ML Chile M. Chapoutier Slightly sweet
Syrah 1500ML German Mastroberardino Sweet
Grenache 360ML Australia Santero Medium dry
Merlot 200ML America Saltarelli Extremely sweet
Tempranillo 3000ML Bordeaux Cavicchioli Medium dry
Sangiovese 1800ML Champagne Fontodi Red Full body
Syrah100% 1000ML Argentina Ca'Rugate Middle sweet
Rakuten has actually over 2000 engineers for IT infrastructure and service development. You know, Rakuten has tons of businesses all over the world, and, we need many many engineers to sustain that. In addition, we have also research institute. That is named “RIT”, standing for Rakuten Institute of Technology. We recruit computer science researchers, mainly phD and are trying to utilize academic knowledge, So as to – For the growth of e-commerce and growth of Rakuten
In order for us to become a global market place, we need to handle items from all over the world and to all over the world, hence we need one uniform database structure. This is the "global catalog" However, there is a serious problem, which have just like political parties, contradicting policies. We have to organize all items globally, but on the other hand we have to adapt locally. In order to solve this problem, we are designing the Hub and spoke model. It has a hub, the item DB in the middle, and a spoke represents an individual local market. For example, "Sports equipment" can be acceptable across different markets, but individual sports, such as baseball or pelota are culture dependent, so they should be treated locally. We are planning to build the first version of the model this summer.
We have applied 2 methods and working on another method. First method is that we have fixed over 2 million noise products from 2 big merchants . This fixing helped them increase their GMS. Second method is that we have fixed over 1 million noise products in 2 large category groups, which are DIY tools and motor cycle categories. These noise corrections also helped GMS to increase 400 million yens according to the number estimated by the marketing team. I'm currently developing another method in Japanese and German Ichiba. My goal is to fix all noise products all over the rakuten world. (--> PREVENTING noise from happening)