rights reserved. In this talk • Overview of Amazon Personalize • How to build Recommenders optimized for Retail and Media & Entertainment • How to identify User Segments • How to use Contextual Information • How to reduce Cold-start Problems • How to optimize Business Metrics
rights reserved. Consumers expect personalized user experiences of consumers see personalization as the standard level of service https://www.business2community.com/marketing/30-amazing-personalization-statistics-02289044
rights reserved. Amazon Personalize Delivers personalized user experiences faster Encrypted to be private and secure Pay only for what you use Responds quickly to changing user intent Easily integrates with existing systems
rights reserved. 추천 알고리즘의 종류 추천 적용 데이터 Hybrid Collaborative Filtering Content-based Filtering Non- Personalization Personalization Cold-Start Problem User-Item Interactions click, purchase, view/watch/listen Content details of articles, reviews, description
rights reserved. 데이터 준비 Users • 사용자 메타 데이터 • 연령, 성별, 고객 멤버십 등 Items • item 메타 데이터 • 가격, SKU(상품 재고 관리 단위), Category, 재고 여부 등 Interactions (required) • 사용자의 item에 대한 행동 로그 데이터 • 구매(buy), 장바구니 담기(cart), 상품 보기(view) 등 CSV 포맷으로 S3에 저장 / 첫번째 rows에 컬럼 Header가 필요함
rights reserved. Solution Version Solution Filter Solution Version Solution version Input Output Batch inference job PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Campaign Filter Bulk Recipe How it works
rights reserved. Recipes Mapped to Use Cases Use Case Recipe User Specific Recommendations User-Personalization User Cold Starts User-Personalization Item Cold Starts User-Personalization Related Items SIMS / Similar-Items Personalized Ranking Personalized-Ranking Popularity Popularity-Count https://github.com/aws-samples/amazon-personalize-samples/blob/master/PersonalizeCheatSheet2.0.md - use-cases-by-recipe
rights reserved. Recipes Mapped to Use Cases Use Case Recipe Media & Entertainment VIDEO_ON_DEMAND Retail ECOMMERCE User Segmentation based on items (e.g., movie, song, product) Item-Affinity User Segmentation based on item attributes (e.g., genres, product categories, brands) Item-Attribute-Affinity New in 2021 https://github.com/aws-samples/amazon-personalize-samples/blob/master/PersonalizeCheatSheet2.0.md - use-cases-by-recipe
rights reserved. Most Popular Because you watched Forrest Gump Use-case-optimized recommenders for media & entertainment • Top picks for you • Because you watched X • More like Y • Most popular Top picks for you VOD Use Cases
rights reserved. Use-case-optimized recommenders for retail • Recommended for you • Customers who viewed this also viewed • Frequently bought together • Most viewed • Best sellers Customers who viewed this also viewed Recommended for You Frequently Bought Together Retail Use Cases
rights reserved. Recommender PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Filter Recipe (VOD/Retail) D O M A I N D A T A S E T G R O U P F O R V O D / R E T A I L U S E C A S E S How it works
rights reserved. Recommender PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Filter Recipe (VOD/Retail) D O M A I N D A T A S E T G R O U P F O R V O D / R E T A I L U S E C A S E S How it works e.g.) • Topic for you - arn:aws:personalize:::recipe/aws-vod-top-picks • Recommended for you - arn:aws:personalize:::recipe/aws-ecomm-recommended-for-you https://docs.aws.amazon.com/personalize/latest/dg/domain-use-cases.html
rights reserved. Recommender PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Filter Recipe (VOD/Retail) D O M A I N D A T A S E T G R O U P F O R V O D / R E T A I L U S E C A S E S How it works
rights reserved. VIDEO_ON_DEMAND Domain Dataset Required for More like X, Top picks for you https://docs.aws.amazon.com/personalize/latest/dg/VIDEO_ON_DEMAND-use-cases.html
rights reserved. Intelligent User Segmentation Identify User segments • Identify users interested in a genre, category, or any other item attribute • Identify users interested in a given item such as a movie, product, etc. • More effective campaigns through marketing channels • Acquire users for new product categories, genres, channels, etc. • Improve return on investment for your marketing spend Action movie fans
rights reserved. User Segmentation Item-Attribute-Affinity: User segmentation based on item attributes Item-Affinity: User segmentation based on items Input Output Amazon Personalize Batch segment job Batch segment job
rights reserved. Recipes Interactions Items Users Item-Affinity Required Optional* Optional* Item-Attribute-Affinity Required Required Optional* R E C I P E T O D A T A S E T T Y P E M A P P I N G F O R T R A I N I N G User Segmentation * Metadata may still be used for filters
rights reserved. Solution Version Solution Filter Solution Version Solution version Input Output Batch segment job PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Engage Item affinity recipe U S E R S E G M E N T A T I O N How it works
rights reserved. https://github.com/aws-samples/amazon-personalize-samples/blob/master/next_steps/core_use_cases/user_segmentation/user_segmentation_example.ipynb User Segmentation Example using Amazon Prime Pantry’s dataset
rights reserved. Contextual recommendations Providing relevant recommendations requires that you consider the context in which they are being viewed Amazon Personalize considers the context while generating recommendations Examples • Device type • Location • Time of day/seasonality
rights reserved. Context (e.g., CABIN_TYPE) https://github.com/aws-samples/amazon-personalize-samples/blob/master/next_steps/core_use_cases/user_personalization/user-personalization-with-contextual- recommendations.ipynb Real-time User Personalization with Contextual Information
rights reserved. Unlock information in unstructured text Valuable signals are often trapped in descriptions, synopses, and reviews Amazon Personalize uses Natural Language Processing (NLP) to automatically extract key information from unstructured text Description: David Attenborough narrates this highly-acclaimed series exploring the natural world of the planet. Each episode explores a different habitat, focusing on how living creatures deal with the challenges posed by each environment. Emmy Award-winning, 11 episodes, 5 years in the making, the most expensive nature documentary ever commissioned by the BBC, and the first to be filmed in high definition. Magnificent. ⭐ ⭐ ⭐ ⭐ ⭐ Amazing. Stunning. Magnificent. Planet Earth goes where no show has gone before. It captures beautiful images and animals that no one has before captured on film. I was absolutely blown away. A True Masterpiece ⭐ ⭐ ⭐ ⭐ Alastair Fothergill's cinematic docu-series was a phenomenon in 2006 - and the touchstone for an extremely consistent franchise of remarkable and evergreen nature documentaries showing us with unparalleled technical ability the majesty and beauty of our planet. Astonishing ⭐ ⭐ ⭐ ⭐ ⭐ The production value is absolutely amazing and is so informative. You honestly can't believe what your watching because it doesn't seem possible. The shots they do are so creative and is well worth the time. A must watch for everyone.
rights reserved. New similar items recipe Similarity algorithms that only look at co-interactions between items are not sufficient New similarity recipe uses item metadata, in addition to co- interactions, to determine similarity More Shows Like First Time Fixer Upper
rights reserved. Recipes Interactions Items Users SIMS Required Ignored for training* Ignored for training* Similar-Items Required Required Ignored for training* R E C I P E T O D A T A S E T T Y P E M A P P I N G F O R T R A I N I N G New similar items recipe * Metadata may still be used for filters
rights reserved. A mix of snacks that are commonly found with a soft drink purchase Similar-Items Comparing similar items’ inference results for a less popular item
rights reserved. Comparing similar items’ inference results for a less popular item SIMS Similar-Items A mix of snacks that are commonly found with a soft drink purchase Falls back to recommending popular items https://github.com/aws-samples/amazon-personalize-samples/blob/master/next_steps/core_use_cases/related_items/personalize_aws_similar_items_example.ipynb
rights reserved. Unstructured Text Best Practices item과 관련성이 높고, 간결하고 명료한 text를 사용 일부 item만 text를 포함한 경우, 오히려 성능 저하가 발생할 수 있음 markup 또는 불필요한 공백(white space, tab 등)을 제거 현재는 English 만 지원 User-Personalization, Personalized-Ranking, Similar-Items 레시피에 적용
rights reserved. New items in fast-changing catalogs (cold start) New items have no interaction history, which makes personalization challenging Amazon Personalize enables you to create a balance between recommendations for new and old items in your catalog Watch as a Family New! New! Item exploration weight 0 1 0.3
rights reserved. Balancing exploration with exploitation Fast moving catalogs & discovery use-cases explorationWeight > 0.5 More stable catalogs & guided use-cases explorationWeight < 0.5 Recommend items with less interaction data & relevance Exploration Exploitation Recommend items based on what we know or relevance explorationWeight = 0.5 https://github.com/aws-samples/amazon-personalize-samples/blob/master/next_steps/core_use_cases/user_personalization/user-personalization-with-exploration.ipynb
rights reserved. Solution Version Solution Filter Solution Version Solution version Input Output Batch inference job PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Campaign Filter Bulk Recipe How it works B A L A N C I N G E X P L O R A T I O N W I T H E X P L O I T A T I O N
rights reserved. Solution Version Solution Filter Solution Version Solution version Input Output Batch inference job PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Campaign Filter Bulk Recipe How it works B A L A N C I N G E X P L O R A T I O N W I T H E X P L O I T A T I O N
rights reserved. Solution Version Solution Filter Solution Version Solution version Input Output Batch inference job PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Campaign Filter Bulk Recipe How it works B A L A N C I N G E X P L O R A T I O N W I T H E X P L O I T A T I O N
rights reserved. Optimize for business metrics Most recommender systems are designed to only increase user engagement Optimize recommendations for relevance while improving business metrics Optimize for revenue, profit margin, video watch time, or any numerical attribute in your catalog
rights reserved. Solution Version Solution Filter Solution Version Solution version Input Output Batch inference job PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Campaign Filter Bulk Recipe O P T I M I Z E F O R B U S I N E S S M E T R I C S How it works
rights reserved. Solution Version Solution Filter Solution Version Solution version Input Output Batch inference job PutEvents click, view/watch/listen, purchase PutItems category, price, genre, description PutUsers age, location, subscription tier Interactions dataset Analytics solution Catalog management system User management system Users dataset Items dataset Campaign Filter Bulk Recipe O P T I M I Z E F O R B U S I N E S S M E T R I C S How it works • HIGH • MEDIUM • LOW • OFF The numerical metadata column from the dataset group's Items dataset that relates to your objective
rights reserved. Minimize costs by recommending movies with lower royalty fees To minimize royalties, we multiply the royalty field by -1 e.g.) 0.05 x (-1) = -0.05 https://github.com/aws-samples/amazon-personalize-samples/blob/master/next_steps/core_use_cases/objective_optimization/objective-optimization.ipynb
rights reserved. The trend of lower royalties as the objective optimization setting is increased from low to high, as you would expect. Offline Metrics Minimize costs by recommending movies with lower royalty fees Recommended movies for the sample user Create the Solution and Version
rights reserved. User-level recommendations Item-item similarity Personalized ranking Use-case-optimized recommendations User segmentation New item bias Business rules/filters Optimize for business metric (e.g., profit, revenue, watch time) Real-time recommendations API Download batch recommendations Contextualize recommendations (e.g., device type, location) User interactions (views, sign-ups, conversions, etc.) Item metadata (details of articles, products, videos, etc.) User metadata (age, location, etc.) Add your data Create a solution Tune recommendations Access recommendations H O W I T W O R K S Amazon Personalize
rights reserved. Interactions Items Users User-Personalization Required Recommended* Optional* SIMS Required Ignored for training* Ignored for training* Similar-Items Required Required Ignored for training* Personalized-Ranking Required Optional* Optional* VIDEO_ON_DEMAND Required Required for More like X, Top picks for you Optional* ECOMMERCE Required Optional* Optional* Item-Affinity Required Optional* Optional* Item-Attribute-Affinity Required Required Optional* * Metadata may still be useful for filters Recipe to Dataset Type Mapping for Training
데이터 준비 모델 배포 Recipes • User-Personalize • SIMS/Similar-Items • Personalized-Ranking • Popularity-Count • more … Datasets • Interactions • Users • Items Solution ㄴ Solution Version • Campaign • Batch Inference Amazon Personalize 추천 성능 개선 방법 모델 개발 모델 학습 Datasets Campaign Solution Recipes Datasets 비즈니스 문제 Batch Inference Datasets
rights reserved. Enhancing customer experiences with Amazon Personalize • Build state-of-the-art recommendation systems • Personalize every customer touchpoint • Engage customers based on their affinity with your products and content • No machine learning experience required • Pay-as-you-go
ksmin23
sudoakiy
kazuhitotakahashi
kaminashi
smt7174
yoshidashingo
tacck
10xinc
tenshoku_draft
ysknsid25
akexorcist
gumamon
shpigford
keithpitt
samanthasiow
pauljervisheath
mraible
bkeepers
kneath
brianwarren
sugarenia
lauravandoore
yeseniaperezcruz
chrisshort
