SAP Inside Track Tokyo 2022 Deep Learning版Cash Applicationをやってみた
1. Dive into the Cash Application
Deep Learning Version
2022-03-10: Week 6 – AI Business Services
TOKYO
SAP Inside Track 2022
2. 2
名前: 小栗俊太郎
会社: SAP
部門: Customer Success
役割: Principal Consultant
SAP Machine Learning Practice Lead
SAP Innovation Service Team Lead
参加目的:情報収集、ネットワーキング(AI/MLな人とつながりたい)
小栗俊太郎
SAP Community
のプロファイルを
画像コピー
6. 8
Intelligent Enterprise
テクノロジー
ANALYTICS APPLICATION
DEVELOPMENT
& INTEGRATION
DATABASE & DATA
MANAGEMENT
INTELLIGENT
TECHNOLOGIES
BUSINESS TECHNOLOGY PLATFORM
ビジネスプロセス
BUSINESS NETWORK
ACROSS ALL FUNCTIONS
BUSINESS PROCESS INTELLIGENCE
アプリケーション
SUSTAINABILITY MANAGEMENT
EXPERIENCE MANAGEMENT
INTELLIGENT SUITE INDUSTRY CLOUD
SAP DATA CENTER HYPERSCALER
インフラ
12. 14
Model Manager for SAP Cash Application
Training Runs Page
NEW
https://help.sap.com/viewer/9668f05979b644cd960f0d83ae4ed0aa/Latest/en-US/a3fd584d0d664f4b854ef7aa1264ac67.html
13. 15
Model Manager for SAP Cash Application
Statistics Page
NEW
https://help.sap.com/viewer/9668f05979b644cd960f0d83ae4ed0aa/Latest/en-US/dda16a97217c4b2d9f227d50a9769857.html
14. 16
Model Manager for SAP Cash Application
Machine Learning Models Page
NEW
https://help.sap.com/viewer/9668f05979b644cd960f0d83ae4ed0aa/Latest/en-US/106b17e7374d4d60b57418d1f0209cd1.html
23. 25
Cash Applicationが解決するべき課題
bank statement fields: numeric, categorical, string
invoice fields: numeric, categorical, string
ASSIGNMENTREFERENCE
BUSINESSPARTNERNAME
DEBITCREDITCODE
DOCUMENTREFERENCEID
MEMOLINE
PAYMENTADVICEACCOUNT
PAYMENTREFERENCE
TRANSACTIONCURRENCY
AMOUNTINTRANSACTIONCURRENCY
............
ACCOUNTINGDOCUMENT_IV
ASSIGNMENTREFERENCE_IV
BILLINGDOCUMENT_IV
COMPANYCODECURRENCY_IV
DEBITCREDITCODE_IV
DEBTOR_IV
DOCUMENTITEMTEXT_IV
DOCUMENTREFERENCEID_IV
ORGANIZATIONBPNAME1_IV
ORGANIZATIONBPNAME2_IV
PAYMENTREFERENCE_IV
TRANSACTIONCURRENCY_IV
AMOUNTINCOMPANYCODECURRENCY_IV
............
bank statement
invoice
single match
multi match
Single Match
Multi Match
24. 26
学習データの構造
Query Key Amount Posting Date Company Code Currency Note Reference ID
Target Key Amount Posting Date Company Code Currency Customer Name Reference ID
1
Reference ID
2
Query Document (QD)
Target Document (TD)
Numeric Fields Categorical Fields Text Fields
Matching Relation (R)
Query Key Target Key
Key Fields
25. 27
Single Match と Multi Match
QD TD
Q1
Q2
QM
T1
T2
T3
TN
R1
R2
R3
RN
R
Single Match
Multi Match
.
.
.
.
.
.
.
.
.
28. 30
深層学習アーキテクチャ概要
Numeric Fields Categorical Fields Text Fields
Character Level
Decomposable Attention
Categorical Embedding
(Trainable)
Feed Forward Network
(1 Residual Connection)
QD TD QD TD
QD TD
Feed Forward Classifier
Single Match Multi Match No Match
In Feedforward Neural Network, data
flows in one direction (from the input
layer to the output layer) between the
layers.
Returns the result of a match of a
numeric field.
Neural architecture for natural language
inference adding sentence attention that is a
method for focusing on important points in the
past (=Attention) when dealing with
continuous data.
Returns the result of a match of a Text
Fields.
Translate large sparse vectors into a
lower-dimensional space that
preserves semantic relationships.
Returns the result of a match of a
Categorical Fields.
Numeric Fields
Categorical Fields Text Fields
Key Fields
33. 35
1. Cash Application概要
2. 既存アルゴリズム
3. Deep Learningの新しいアルゴリズム
4. 実証結果
Deep Learning版Cash Applicationをやってみた
34. 36
◼ ネガティブサンプリングを含む深層学習を行った結果、 Single Matchの精度90~95%台と
大きく変わらないが、Multi Matchの精度が飛躍的に向上した。
◼ 例 : (会社コードのデータに依存)
◼ Multi Match Proposal rate : 24% -> 38%
◼ Multi Match Accuracy Rate :60% ->87.5%
◼ Multi Match Proposal rate : 0% -> 61.54%
◼ Multi Match Accuracy Rate :0% ->95.83%
検証結果
37. 39
参照1: Model Manager for SAP Cash Application
▪ https://help.sap.com/viewer/9668f05979b644cd960f0d83ae4ed0aa/Latest/en-US/73fd86f7258b46bf84ef47f415af18fa.html
参照2: Entity Embeddings of Categorical Variables
▪ https://arxiv.org/abs/1604.06737
参照3: A Decomposable Attention Model for Natural Language Inference
▪ https://arxiv.org/abs/1606.01933
参照4: Papers With Code: The latest in Machine Learning
▪ https://paperswithcode.com/
参照5: Switch to Deep Learning Line-Item Matching
▪ https://help.sap.com/viewer/60e4f6eb9b6a427db1963afb7e9791b5/Latest/en-
US/5eaba15e12044f299efc95d8aa236154.html
引用、参考、参照、転載の一覧