Actual cases of applying AI related technologies
in Rakuten
Yu Hirate, Dr. Eng.
Rakuten Institute of Technology Tokyo,
Rakuten, Inc.
2
Yu Hirate
 Principal Scientist, Rakuten Institute of Technology
Manager, Intelligence domain research group
 Bio.
• 2005-2008 CS div. graduate school of Science and Engineering, Waseda
University.
• 2006-2009 Research Associate, Media Network Center, Waseda University.
• 2009- current Rakuten Institute of Technology.
Working on projects for extracting knowledge from large scale of data
by utilizing data mining, machine learning technologies.
3
Who is Rakuten?
4
Rakuten Institute of Technology
Masaya Mori
Global head
• Established in 2006.
• Launched R.I.T. NY in 2010.
• Launched R.I.T. Paris in 2014.
• Launched R.I.T. Singapore / Boston in 2015.
Strategic R&D organization for Rakuten Group
5
Rakuten Institute of Technology
+100 researchers in 5 locations
Tokyo NYSingapore
Paris
Boston
6
Research Groups in RIT
3 research groups for adapting with Internet growth
RealityIntelligencePower
• HCI
• AR / VR
• Image Processing
• Distributed Computing
• HPC
• IoT
• Machine Learning
• Deep Learning
• NLP
• Data Mining
7
Mission of Intelligence Domain Group
Assisting Rakuten businesses to boost with AI and Data Mining
Optimizing A/B testing
Item Classification
User Segmentation
AI
Coupon Distribution
Recommender System Economy Prediction / Demand Prediction
Review Analysis
Anomaly Detection / Fraud Detection
Image Recognition
8
1. Product Data Science
9
Product Data in Rakuten
Huge!
Unstructured!
241 million items
Num.ofItems(million)
date
10
Product Data Science
Product Catalog
Organize Large Amounts of Unstructured Information Into Structured Knowledge
Why are we working on this problem? (Key Benefits)
‣ To organize our catalog in accordance with customer
expectations
‣ To precisely search our catalog for products and its variants
‣ To measure and enforce merchant KPI's.
What are we doing? (Key Tasks)
‣ Product Genre Classification
‣ Attribute Extraction from Product Information
‣ Merchant and Item Review Analysis
How are we doing? (Key Technologies)
‣ Large-Scale Gradient Boosted Decision Trees
‣ Deep Learning (RNN's, CNN's, others)
‣ Computing Massive Number of NLP Features
Businesses
11
Generating Structured Product Data
Each product can be assigned a category and attributes. For instance:
+Category Grocery & food
Subcategory Wine
Each (sub)category has a number of relevant attributes with a list of valid values
Challenge: this structured information is not always present or correct
Goal: automatically predict category and attributes from text and/or images
12
Product Category Classification
13
Category Classification based on Deep Learning
Classifier based on
Deep Learning Algorithm (CNN)
Prec@1 92%
Prec@10 99%
Classifier based on
Deep Learning Algorithm(CNN)
Prec@1 57%
Prec@3 75%
Extracting Words
* Tested to Ichiba L3 category (1.5K categories)
* Tested for PriceMinister Image Data
Text Data
• Item Title
• Item Description
Image Data
14
Category Classification from Image Data
Hobby and Entertainment
> Books and Magazine
> Business Electronics
> Audio
> Earphone / Headphone
Electronics
> Smartphone
> AC Adaptor / Battery
15
16
Attribute Extraction from Text and Images
Text : Automatic Rule Generation
with Machine Learning
Images:
Attributes prediction with Deep Learning.
attribute top-1 accuracy # of values
type 74% 94
material 66% 84
shape 83% 5
color 75% 24
PriceMinister furniture category
17
Review Analysis : Informative Item Review Extraction
Informative
Positive Review
Informative
Negative Review
Non-informative
Informative
PositiveNegative
Classify All Item Reviews with two perspective
amount of information and polarity.
Pick up two reviews
Boosted up Ichiba GMS by 15 billion JPY
(as of 2015)
18
Review Analysis : Merchant Review Analysis
Simple moving average
Reviews corresponding to the category
(evidence is highlighted)
category
Positive/negative
19
Review Analysis : Travel Review Analysis
in collaboration with
Our Review Analysis Algorithm has
also been applied to Travel breakfast
festival
20
2. User Behavior Analysis
21
User Behavior Data
Purchase History Data Search Query Log Data
date date
DailyTransactionCount
DailySearchRequests
22
Utilizing User Behavior Analysis
Extracting user preference, and
applying to various functions.
Discount CouponItem Recommender System
Ads
Optimization
Prospective
User Extraction
23
1. Recommender Algorithm based on Distributed Representation
CategoriesMerchants
Users
Items
Services
Similar User
Item Recommender
Coupon Item
Recommender
Category
Prediction
Genre
Prediction
Related Categories
→ Navigations
Category
Navigation
Merchant
Recommender
Similar shops
(competitor)
Related Shops
Transforming {items, users, services, categories}
to distributed representation (vectors), and computing similarities.
24
Recommender Algorithm based on Distributed Representation
Utilizing word2vec, doc2vec algorithms
• Similar words are projected into similar vectors.
• Relationship between words can be expressed as a simple
vector calculation.
• Analogy
• v(“woman”) – v(“man”) + v(“king”) = v(“queen”)
[T.Mikolov et al. NIPS 2013]
25
Sample results of word2vec
query: coffee
• cocoa 0.603515
• robusta 0.565269
• beans 0.565232
• bananas 0.565207
• cinnamon 0.556771
• citrus 0.547495
• espresso 0.542120
• caff 0.542082
• infusions 0.538069
• tea 0.532565
Example 2
(CalTech-101)
Example 1
(English Wikipedia)
26
Applying doc2vec algorithm to recommender system
Document : a sequence of words with context.
User : a sequence of item views with user’s intention.
Set of documents
Vectors for words
Vectors for documents
sim{word, word}
sim{doc, word}
sim{doc, doc}
Set of user behaviors
Vectors for items
Vectors for users
sim{item, item}
sim{user, item}
sim{user, user}
27
Sample Results
Query
RecommendedResults
28
Performance Evaluation
3.42
8.02
18.2
21.04
24.17
0
5
10
15
20
25
30
Item Similarity Matrix
Factorization
Doc2Vec-based Item Vector-
based (from
Word2vec)
Item Vector-
based (from
Doc2vec)
hit-rate(%)
Distributed Representation Based AlgorithmConventional methods
Hit-Rate : Percentage of Users who bought recommended items.
29
Applying actual functions
Discount Coupon
Item Recommender System
Recommender algorithms w/ distributed representations are
testing in the actual services in Ichiba.
30
2. Extracting Potential Customers
• We are providing 70+ services in
Japan that covers human life.
• Each service has its on purchase
history data, browsing history data.
• Those data are connected with a
single ID, Rakuten ID.
31
Extracting Potential Customers
Detect prospective applicants from Ichiba purchasers
by using their purchase trends and demographics
Ichiba Active
Users
Prospective
Applicants
Extract a Fintech service
32
Extracting Potential Customers : Approach
Ichiba
Active Users
Overlap
7,413
Positive Samples
7,417
Negative Samples
About 50% of contractors of the Fintech service were Ichiba Active Users.
33
Extracting Potential Customers : Significant Factors in the customer
model
0 0.1 0.2 0.3
genre_41_100890_ / 花・ガーデン・DIY / DIY・工具
genre_72_111078_ / キッズ・ベビー・マタニティ / キッズ
genre_50_110983_ / 靴 / メンズ靴
Age-05-[35-40]
genre_93_101077_ / スポーツ・アウトドア / ゴルフ
Area-01-Kanto
Area-00-Others
genre_113_101126_ / 車用品・バイク用品 / カー用品
Age-08-[50-*]
Age-03-[25-30]
Age-00-none
gms
Gender-00-none
basket_max_price
frequency
basket_average_price
average_unit_price
Age-02-[20-25]
Gender-02-female
Gender-01-male
Top 20 factors selected from 141 factors
Ichiba Genre /Cars & Motorcycles/Car Accessories
Ichiba Genre /Sports & Outdoors/Golf
Ichiba Genre /Shoes/Men’s Shoes
Ichiba Genre /Kids & Baby/Kids
Ichiba Genre /Gardening & Tools / DIY Tools
average_unit_price
basket_average_price
frequency
basket_max_price
gms
Loyalty
Life Stage
34
Extracting Potential Customers : Evaluation for the model
Prospective Users Control Group
• Randomly Selected
• About 300,000 users
• Score >= 0.8
• About 300,000 users
Send ichiba mail magazine to two groups
Ichiba Mail Magazine
35
Extracting Potential Customers : Evaluation Result
Mail Deliver
Open Mail
Click Contents
(Visit Service
Page)
Click Rate went up by +49.23%
compared with control group
+3.52% +49.23%
36
Actual Cases in Rakuten
Premium Membership Service in Japan Ichiba (E-Commerce Service)
Existing Customers
(~31K users)
Ichiba
Purchase
History Data
User
Demographic
Data
Data used for Customer Model
Conducted Email Marketing to Japan Ichiba Users
RIT
Customer
Model
Normal
Comparison of CVR
+123.30%
37
Other Cases
Potential Home Loan Contractors for Rakuten Bank
Potential mobile service subscribers for Rakuten Mobile.
Potential contents purchasers for showtime (online video streaming service).
Potential Life Insurance contractors for Rakuten Insurance
Advertisement Targeting
Potential customers for offline shops
38
3. Geolocation Based Marketing
Understand user’s preference
>> make appropriate recommendations
Locate users & Explore their behavior patterns
>> provide other options
>> expand usage coverage
>> extract potential cross-service users
>> facilitate cross-use between services
Rakuten Strategy: Fusion of the Internet and the Offline
39
Location Based Recommender System
Which shops should be offered
to users located at this point?
User Viewpoint Shop Viewpoint
Who are potential customers
for this shop?
Issue : How to Identify major areas for each users?
40
Location Based Recommender System
Centroid
Shop coordinates
Boundary
Step 1 : scatter used shops
Step 4 : locate recommended shop
Step 2 : label centroid, build boundary
Step 3 : extract all shops
41
Effectiveness of geolocation based recommender
CVR (%) gms/user (JPY)
×8.6
×1.3
Send emails to make recommendations to users within their major areas
Both CVR and GMS/uu are improved
Geolocation
based
recommender
Geolocation
based
recommender
42
4. Others
43
Janken
44
Logistics Pick Path Optimization
我们真的很有诚意了。
你说我一个老总都亲
自跑了好几趟了。
Speech
recognition
Machine
translation
45
is a Rakuten group company which provides video streaming service.
Volunteers are editing subtitles and translated subtitles.
Machine Translations for TV dramas
 Translate from Chinese to English sentences
 Extracted 10,000 Chinese-English sentence-pairs to
evaluate commercial APIs and IBot, e.g.,
 我一个老总都亲自跑了好几趟了
 I’m a director and yet I’ve made so many trips
 Extracted another 2.1 million sentence-pairs to train
IBot’s model
46
Data Preparation
 Applying Attentional Recurrent Neural Networks
(RNN)
 Neural Machine Translation by Jointly Learning to
Align and Translate [Bahdanau, Cho & Bengio, ICLR 2015]
 658 citations (Google scholar)
 Train RNN with 2.1 million c
Chinese-English sentence
pairs
47
RIT’s deep learning based MT model
 Evaluated on 10,000 Chinese-English sentence pairs
System BLEU (%) METEOR (%)
Google API 12 20
Microsoft API 12 20
IBM Watson API 3 12
RIT (Aug 24) 10 15
RIT (Sep 7) 14 19
RIT (Sep 21) 22 24
RIT (Nov 28) 36 30
48
RIT MT module outperforms in TV drama domain
49

Actual cases of applying AI related technologiesin Rakuten

  • 1.
    Actual cases ofapplying AI related technologies in Rakuten Yu Hirate, Dr. Eng. Rakuten Institute of Technology Tokyo, Rakuten, Inc.
  • 2.
    2 Yu Hirate  PrincipalScientist, Rakuten Institute of Technology Manager, Intelligence domain research group  Bio. • 2005-2008 CS div. graduate school of Science and Engineering, Waseda University. • 2006-2009 Research Associate, Media Network Center, Waseda University. • 2009- current Rakuten Institute of Technology. Working on projects for extracting knowledge from large scale of data by utilizing data mining, machine learning technologies.
  • 3.
  • 4.
    4 Rakuten Institute ofTechnology Masaya Mori Global head • Established in 2006. • Launched R.I.T. NY in 2010. • Launched R.I.T. Paris in 2014. • Launched R.I.T. Singapore / Boston in 2015. Strategic R&D organization for Rakuten Group
  • 5.
    5 Rakuten Institute ofTechnology +100 researchers in 5 locations Tokyo NYSingapore Paris Boston
  • 6.
    6 Research Groups inRIT 3 research groups for adapting with Internet growth RealityIntelligencePower • HCI • AR / VR • Image Processing • Distributed Computing • HPC • IoT • Machine Learning • Deep Learning • NLP • Data Mining
  • 7.
    7 Mission of IntelligenceDomain Group Assisting Rakuten businesses to boost with AI and Data Mining Optimizing A/B testing Item Classification User Segmentation AI Coupon Distribution Recommender System Economy Prediction / Demand Prediction Review Analysis Anomaly Detection / Fraud Detection Image Recognition
  • 8.
  • 9.
    9 Product Data inRakuten Huge! Unstructured! 241 million items Num.ofItems(million) date
  • 10.
    10 Product Data Science ProductCatalog Organize Large Amounts of Unstructured Information Into Structured Knowledge Why are we working on this problem? (Key Benefits) ‣ To organize our catalog in accordance with customer expectations ‣ To precisely search our catalog for products and its variants ‣ To measure and enforce merchant KPI's. What are we doing? (Key Tasks) ‣ Product Genre Classification ‣ Attribute Extraction from Product Information ‣ Merchant and Item Review Analysis How are we doing? (Key Technologies) ‣ Large-Scale Gradient Boosted Decision Trees ‣ Deep Learning (RNN's, CNN's, others) ‣ Computing Massive Number of NLP Features Businesses
  • 11.
    11 Generating Structured ProductData Each product can be assigned a category and attributes. For instance: +Category Grocery & food Subcategory Wine Each (sub)category has a number of relevant attributes with a list of valid values Challenge: this structured information is not always present or correct Goal: automatically predict category and attributes from text and/or images
  • 12.
  • 13.
    13 Category Classification basedon Deep Learning Classifier based on Deep Learning Algorithm (CNN) Prec@1 92% Prec@10 99% Classifier based on Deep Learning Algorithm(CNN) Prec@1 57% Prec@3 75% Extracting Words * Tested to Ichiba L3 category (1.5K categories) * Tested for PriceMinister Image Data Text Data • Item Title • Item Description Image Data
  • 14.
    14 Category Classification fromImage Data Hobby and Entertainment > Books and Magazine > Business Electronics > Audio > Earphone / Headphone Electronics > Smartphone > AC Adaptor / Battery
  • 15.
  • 16.
    16 Attribute Extraction fromText and Images Text : Automatic Rule Generation with Machine Learning Images: Attributes prediction with Deep Learning. attribute top-1 accuracy # of values type 74% 94 material 66% 84 shape 83% 5 color 75% 24 PriceMinister furniture category
  • 17.
    17 Review Analysis :Informative Item Review Extraction Informative Positive Review Informative Negative Review Non-informative Informative PositiveNegative Classify All Item Reviews with two perspective amount of information and polarity. Pick up two reviews Boosted up Ichiba GMS by 15 billion JPY (as of 2015)
  • 18.
    18 Review Analysis :Merchant Review Analysis Simple moving average Reviews corresponding to the category (evidence is highlighted) category Positive/negative
  • 19.
    19 Review Analysis :Travel Review Analysis in collaboration with Our Review Analysis Algorithm has also been applied to Travel breakfast festival
  • 20.
  • 21.
    21 User Behavior Data PurchaseHistory Data Search Query Log Data date date DailyTransactionCount DailySearchRequests
  • 22.
    22 Utilizing User BehaviorAnalysis Extracting user preference, and applying to various functions. Discount CouponItem Recommender System Ads Optimization Prospective User Extraction
  • 23.
    23 1. Recommender Algorithmbased on Distributed Representation CategoriesMerchants Users Items Services Similar User Item Recommender Coupon Item Recommender Category Prediction Genre Prediction Related Categories → Navigations Category Navigation Merchant Recommender Similar shops (competitor) Related Shops Transforming {items, users, services, categories} to distributed representation (vectors), and computing similarities.
  • 24.
    24 Recommender Algorithm basedon Distributed Representation Utilizing word2vec, doc2vec algorithms • Similar words are projected into similar vectors. • Relationship between words can be expressed as a simple vector calculation. • Analogy • v(“woman”) – v(“man”) + v(“king”) = v(“queen”) [T.Mikolov et al. NIPS 2013]
  • 25.
    25 Sample results ofword2vec query: coffee • cocoa 0.603515 • robusta 0.565269 • beans 0.565232 • bananas 0.565207 • cinnamon 0.556771 • citrus 0.547495 • espresso 0.542120 • caff 0.542082 • infusions 0.538069 • tea 0.532565 Example 2 (CalTech-101) Example 1 (English Wikipedia)
  • 26.
    26 Applying doc2vec algorithmto recommender system Document : a sequence of words with context. User : a sequence of item views with user’s intention. Set of documents Vectors for words Vectors for documents sim{word, word} sim{doc, word} sim{doc, doc} Set of user behaviors Vectors for items Vectors for users sim{item, item} sim{user, item} sim{user, user}
  • 27.
  • 28.
    28 Performance Evaluation 3.42 8.02 18.2 21.04 24.17 0 5 10 15 20 25 30 Item SimilarityMatrix Factorization Doc2Vec-based Item Vector- based (from Word2vec) Item Vector- based (from Doc2vec) hit-rate(%) Distributed Representation Based AlgorithmConventional methods Hit-Rate : Percentage of Users who bought recommended items.
  • 29.
    29 Applying actual functions DiscountCoupon Item Recommender System Recommender algorithms w/ distributed representations are testing in the actual services in Ichiba.
  • 30.
    30 2. Extracting PotentialCustomers • We are providing 70+ services in Japan that covers human life. • Each service has its on purchase history data, browsing history data. • Those data are connected with a single ID, Rakuten ID.
  • 31.
    31 Extracting Potential Customers Detectprospective applicants from Ichiba purchasers by using their purchase trends and demographics Ichiba Active Users Prospective Applicants Extract a Fintech service
  • 32.
    32 Extracting Potential Customers: Approach Ichiba Active Users Overlap 7,413 Positive Samples 7,417 Negative Samples About 50% of contractors of the Fintech service were Ichiba Active Users.
  • 33.
    33 Extracting Potential Customers: Significant Factors in the customer model 0 0.1 0.2 0.3 genre_41_100890_ / 花・ガーデン・DIY / DIY・工具 genre_72_111078_ / キッズ・ベビー・マタニティ / キッズ genre_50_110983_ / 靴 / メンズ靴 Age-05-[35-40] genre_93_101077_ / スポーツ・アウトドア / ゴルフ Area-01-Kanto Area-00-Others genre_113_101126_ / 車用品・バイク用品 / カー用品 Age-08-[50-*] Age-03-[25-30] Age-00-none gms Gender-00-none basket_max_price frequency basket_average_price average_unit_price Age-02-[20-25] Gender-02-female Gender-01-male Top 20 factors selected from 141 factors Ichiba Genre /Cars & Motorcycles/Car Accessories Ichiba Genre /Sports & Outdoors/Golf Ichiba Genre /Shoes/Men’s Shoes Ichiba Genre /Kids & Baby/Kids Ichiba Genre /Gardening & Tools / DIY Tools average_unit_price basket_average_price frequency basket_max_price gms Loyalty Life Stage
  • 34.
    34 Extracting Potential Customers: Evaluation for the model Prospective Users Control Group • Randomly Selected • About 300,000 users • Score >= 0.8 • About 300,000 users Send ichiba mail magazine to two groups Ichiba Mail Magazine
  • 35.
    35 Extracting Potential Customers: Evaluation Result Mail Deliver Open Mail Click Contents (Visit Service Page) Click Rate went up by +49.23% compared with control group +3.52% +49.23%
  • 36.
    36 Actual Cases inRakuten Premium Membership Service in Japan Ichiba (E-Commerce Service) Existing Customers (~31K users) Ichiba Purchase History Data User Demographic Data Data used for Customer Model Conducted Email Marketing to Japan Ichiba Users RIT Customer Model Normal Comparison of CVR +123.30%
  • 37.
    37 Other Cases Potential HomeLoan Contractors for Rakuten Bank Potential mobile service subscribers for Rakuten Mobile. Potential contents purchasers for showtime (online video streaming service). Potential Life Insurance contractors for Rakuten Insurance Advertisement Targeting Potential customers for offline shops
  • 38.
    38 3. Geolocation BasedMarketing Understand user’s preference >> make appropriate recommendations Locate users & Explore their behavior patterns >> provide other options >> expand usage coverage >> extract potential cross-service users >> facilitate cross-use between services Rakuten Strategy: Fusion of the Internet and the Offline
  • 39.
    39 Location Based RecommenderSystem Which shops should be offered to users located at this point? User Viewpoint Shop Viewpoint Who are potential customers for this shop? Issue : How to Identify major areas for each users?
  • 40.
    40 Location Based RecommenderSystem Centroid Shop coordinates Boundary Step 1 : scatter used shops Step 4 : locate recommended shop Step 2 : label centroid, build boundary Step 3 : extract all shops
  • 41.
    41 Effectiveness of geolocationbased recommender CVR (%) gms/user (JPY) ×8.6 ×1.3 Send emails to make recommendations to users within their major areas Both CVR and GMS/uu are improved Geolocation based recommender Geolocation based recommender
  • 42.
  • 43.
  • 44.
  • 45.
    我们真的很有诚意了。 你说我一个老总都亲 自跑了好几趟了。 Speech recognition Machine translation 45 is a Rakutengroup company which provides video streaming service. Volunteers are editing subtitles and translated subtitles. Machine Translations for TV dramas
  • 46.
     Translate fromChinese to English sentences  Extracted 10,000 Chinese-English sentence-pairs to evaluate commercial APIs and IBot, e.g.,  我一个老总都亲自跑了好几趟了  I’m a director and yet I’ve made so many trips  Extracted another 2.1 million sentence-pairs to train IBot’s model 46 Data Preparation
  • 47.
     Applying AttentionalRecurrent Neural Networks (RNN)  Neural Machine Translation by Jointly Learning to Align and Translate [Bahdanau, Cho & Bengio, ICLR 2015]  658 citations (Google scholar)  Train RNN with 2.1 million c Chinese-English sentence pairs 47 RIT’s deep learning based MT model
  • 48.
     Evaluated on10,000 Chinese-English sentence pairs System BLEU (%) METEOR (%) Google API 12 20 Microsoft API 12 20 IBM Watson API 3 12 RIT (Aug 24) 10 15 RIT (Sep 7) 14 19 RIT (Sep 21) 22 24 RIT (Nov 28) 36 30 48 RIT MT module outperforms in TV drama domain
  • 49.