Contenu connexe Similaire à 아마존닷컴처럼 Amazon Forecast로 시계열 예측하기 - 강지양 솔루션즈 아키텍트, AWS / 강태욱 매니저, GSSHOP :: AWS Summit Seoul 2019 (20) Plus de Amazon Web Services Korea (20) 아마존닷컴처럼 Amazon Forecast로 시계열 예측하기 - 강지양 솔루션즈 아키텍트, AWS / 강태욱 매니저, GSSHOP :: AWS Summit Seoul 20191. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
아마존닷컴처럼 Amazon
Forecast로 시계열 예측하기
강지양
솔루션즈아키텍트
AWS
S e s s i o n I D
강태욱
매니저
GSSHOP
2. © 2019, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
© 2019, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
Amazon Forecast 소개
시계열 예측에 대해 궁금한 것들
적용 사례: GSSHOP
모범 사례
목차
3. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
반복되는 고객의 요청...
“아마존닷컴의 머신 러닝 기술을
우리가 사용할 방법이 있을까요?”
4. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Time-Series Forecasting
시계열 예측
상품
수요 예측
재무지표
예측
노동력
수요 예측
재고 관리
예측
5. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
외부 요인
날씨, 휴일, 이벤트,
유행 트렌드를 반영하려면
과거를 모르는 데이터
이전 데이터가 없는 새로운 제품을 출시했다면
기타 변수
다양한 제품 속성과 관련 데이터를 고려하려면
튀거나 간헐적인 데이터
불규칙한 패턴의 현실 데이터를 예측한다면
정확한 예측은 매우 어렵습니다
6. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
기존의 문제 접근 방식
Traditional
ARIMA
Deep Learning
RNN
Regression
XGBoost
하이퍼파라미터
최적화 문제
메타데이터
반영 문제
데이터의 연속성
반영 문제
7. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon Forecast
Amazon.com의 기술로 만든
정확도 높은 시계열 예측 서비스
ML 전문가가 아니라도
문제 없습니다.
8. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
아마존닷컴의 기술로 어려웠던 예측 문제를 해결
급격한 변화
예측
이전 데이터가
없는 새로운
제품의 수요 예측
연관된 시계열
데이터들간의
관계 학습
외부 요인 반영
(계절, 공휴일,
할인 행사 등)
9. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
데이터
파이프라인
구성
모델 배포
다섯번 클릭으로 만드는 커스텀 딥러닝 모델
모델들의
정확도 메트릭
비교
AutoML 또는
Amazon Forecast
제공 알고리즘
중에서 선택
예측 실행
10. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon Forecast의 동작 방식
Amazon Forecast
11. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon Forecast Predefined Recipes
The DeepAR+ Recipe
The Multi-Quantile Recurrent Neural Network
(MQRNN) Recipe (NIPS 2017)
The Spline Quantile Forecaster (SQF) Recipe
The Non-Parametric Time Series (NPTS) Recipe
The Mixture Density Networks (MDN) Recipe
The ARIMA Recipe
The Exponential Smoothing (ETS) Recipe
The Prophet Recipe
머신러닝 경험 없이도 예측 시스템을 만들 수 있도록 미리 만들어진 레시피를 제공합니다.
12. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
다양한 예측 문제에 적용
13. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
14. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
한눈에 보는 시계열 예측
15. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
한눈에 보는 시계열 예측
16. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
17. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
https://en.wikipedia.org/wiki/Probabilistic_forecasting
18. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
확률 예측이란?
(μ, σ)
19. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
중요한 이유
20. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
21. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
22. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
확률 예측의 결과값
P10, P50, P90: “Quantiles”
23. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
강태욱
매니저
GSSHOP
24. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
강태욱 매니저 | 물류본부
25. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
디지털 트랜스포메이션
26. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
DevOps
AWS Cloud
Hybrid
Cloud
Micro Service
Architecture
IDC
Monolithic
Optimized
27. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
배송주문 물류센터 물류 관점
현황/지표
Data Driven Decision Making
28. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
물동량 예측
물류/배송 효율화
Amazon
SageMaker
Amazon
Forecast
29. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon Forecast (Preview)
Predefined Dataset Domain
Domain Dataset
Recipes
시작하기 – 공식 문서, 권한 설정, Github 샘플 데이터 & 코드
PoC
결론
30. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Predefined Dataset Domain
31. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Retail Domain 데이터 준비
TV쇼핑 상품의 2013년 ~ 현재
• TARGET_TIME_SERIES : 상품아이디, TIMESTAMP, 출고량
• RELATED_TIME_SERIES : 상품아이디, TIMESTAMP, 방송여부 (0, 1)
• ITEM_METADATA Dataset : 상품아이디, TIMESTAMP, 브랜드, 색상
32. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
RELATED_TIME_SERIES DATA (retail domain)
HolidayCalendarName
33. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Recipes
34. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
PoC
35. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
PoC
36. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
SDK
S3
Amazon
Forecast
Aurora
PoC 시스템
recipe
train
model
EC2물류대시보드
37. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
POC 결과
38. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
POC 결과
39. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
POC 결과
40. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon Forecast 결론
8 hours
41. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
42. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
모범 사례
빅데이터의 힘을 믿으세요
데이터가 많을수록 정확합니다
모델을 너무 나누지 마세요
관련 있는 시계열 데이터들의 공통 모델을 만드세요
TARGET 데이터로 시작하세요
보조 데이터는 주인공이 아닙니다
시계열 데이터를 자르지 마세요
끊기지 않은 전체 데이터가 필요합니다
43. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
데이터가 많을수록 정확합니다
Scaling to Very Very Large Corpora for Natural Language
Disambiguation, Banko and Brill, Microsoft Research (2001)
http://www.aclweb.org/anthology/P01-1005
“These results suggest that we may
want to reconsider the trade-off
between spending time and money
on algorithm development versus
spending it on corpus development.”
알고리즘 선정보다
학습 데이터의 양이 더 중요
44. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
45. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
데이터
머신러닝
도메인
지식
균형
46. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon Forecast
개발자도 쉽게 접근할 수 있는 시계열 예측
주 요 기 능
높은 정확도 자동 최적화 통합 모델링콜드 스타트 자동 재학습
47. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Next Step
• Amazon Forecast (Preview)
https://aws.amazon.com/forecast/
• Amazon Forecast Documentation (Preview)
https://docs.aws.amazon.com/forecast/
• 실습 URL (Jupyter Notebook)
https://github.com/aws-samples/amazon-forecast-samples
• AWS re:Invent 2018: [NEW LAUNCH!] Introducing Amazon Forecast (AIM344)
https://youtu.be/kU5cFdXjwOc
• AWS Online Tech Talks: Build Forecasts and Individualized Recommendations with AI
https://youtu.be/glSFmuAfRjE