[NEW LAUNCH!] Introducing Amazon Personalize: Real-time Personalization and Recommendations (AIM365) - AWS re:Invent 2018
•Als PPTX, PDF herunterladen•
9 gefällt mir•4,564 views
Amazon Personalize is a fully-managed service that helps companies deliver personalized experiences, such as recommendations, search results, email campaigns and notifications. It brings over 20 years of experience in personalization from Amazon.com and puts it in the hands of developers with little or no machine learning experience. Amazon Personalize uses AutoML to automate the entire process of managing and processing data, choosing the right algorithm based on the data, and using the data to train and deploy custom machine learning models — all with a few simple API calls. Join us and learn how you can use Concierge to build engaging experiences that respond to user preferences and behavior in real-time.
Empfohlen
Empfohlen
Weitere ähnliche Inhalte
Was ist angesagt?
Was ist angesagt? (20)
Ähnlich wie [NEW LAUNCH!] Introducing Amazon Personalize: Real-time Personalization and Recommendations (AIM365) - AWS re:Invent 2018
Ähnlich wie [NEW LAUNCH!] Introducing Amazon Personalize: Real-time Personalization and Recommendations (AIM365) - AWS re:Invent 2018 (20)
Mehr von Amazon Web Services
Mehr von Amazon Web Services (20)
[NEW LAUNCH!] Introducing Amazon Personalize: Real-time Personalization and Recommendations (AIM365) - AWS re:Invent 2018
- 1. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. IntroducingAmazon Personalize (Preview) Bindu Reddy GM, AWS AI A I M 3 6 5
- 2. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Agenda Overview of service Science of personalization Customer presentations • Domino’s Pizza Enterprises • RB Media
- 3. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Personalizing user experience is proven to increase discoverability, engagement, user satisfaction, and revenue 30% of page views on Amazon are from recommendations … However, most customers find personalization hard to get right
- 4. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Effective personalization requires solving multiple hard problems Reacting to user interactions in real time Avoiding mostly showing popular items Handling cold start (insufficient data about new users/items) Scale
- 5. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Traditional recommender systems aren’t adequate • Rule-based systems perform poorly, don’t scale and are hard to maintain • Collaborative filtering and matrix factorization methods are good for v1, but deep neural networks, esp. recurrent neural networks, that take into account the sequence of a user’s activity (clicks) out-perform other methods
- 6. Deep learning techniques have a direct impact on the bottom line SimilarityPopularity Neural network Matrix factorization +15.4% Engagement Recurrent Neural Net + Bandit Rule-based card ranker Bayesian network model +7.4% Engagement+29% Click Through +20% Click Through
- 7. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Introducing Amazon Personalize Real-time personalization and recommendation, based on the machine learning technology used by Amazon.com Generates personalized recommendations for your users via API calls • State-of-the-art deep learning algorithms that out-perform traditional methods • Auto-ML capabilities that automate the entire process from data ingestion to inference • Real-time personalized recommendations – Ability to ingest activity/clickstream data and generate recommendations, in real time, based on the session/context of the use
- 8. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Common applications & usecases Personalized recommendations Search reranking Notifications and emailsRelated Items
- 9. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Applicable across multipledomains Amazon Personalize can be applied to many domains including • Retail and E-commerce • Video on demand • News • Travel • Personalized notifications
- 10. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Personalize ‒Overview Prepare your data, then upload with the Amazon Personalize API Choose one of our algorithms or tell AutoML to find the best fit Modify your code Retrain continually to improve the model
- 11. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Real-timedatacan beconsumed byAmazon Personalize Historical user activity User attributes Item catalog Real-time data Mobile SDKs (coming soon) JavaScript SDK Amazon S3 bucket Server-Side SDKs Offline data
- 12. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. UsetheConsole/API totrainand experiment withmodels Use AutoML or pick a pre- defined algorithm
- 13. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Deploy thebestmodels (solutions)bylaunching a campaign Amazon Personalize Campaigns • Launching a campaign will deploy all the infrastructure needed to create a personalize endpoint • You can use a simple API to getRecommendatons() or getPersonalizedResults()
- 14. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Deliverpersonalized recommendations inproduction Your web server Browser Real-time event data
- 15. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Consoleto monitor and track allthedifferent steps
- 16. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Easytouseconsoletotestpersonalizedrecommendations
- 17. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon PersonalizeCustomers
- 18. "At Sony, we are focused on developing predictive models to create an engaging and relevant experience for all our consumers. However, delivering truly personalized offers and tailored content to millions of users across channels and content formats is no simple task – it requires the complex integration of advanced technology, people, and processes. Amazon Personalize eliminates the heavy lifting of building a personalized recommendation system. We can leverage Amazon Personalize's data pipelines and indexing infrastructure, while developing our models using Amazon SageMaker. The combination of Amazon SageMaker and Amazon Personalize enables us to automate and accelerate our machine learning development, and drive more effective personalization at scale." Gabor Melli Senior Director of Machine Learning, Sony
- 19. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Balakrishnan (Murali) Narayanaswamy Senior Machine Learning Scientist Amazon AI
- 20. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Modeling for Personalization (order and timing matters) Fundamental insight The evolution of historical interest and dis-interest is a good indicator of future preferences (application data-specific patterns)
- 21. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Modeling for Personalization (order and timing matters) Fundamental insight The evolution of historical interest and dis-interest is a good indicator of future preferences (application data specific patterns)
- 22. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Sequentialmodeling (recurrent neural network) (labels) Available in Amazon Personalize as Recurrent Recommender (features) (learned user representation) Other Information User Representation
- 23. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. ModelingSessions Available in Amazon Personalize as Hierarchical Recurrent Neural Network (HRNN) (learned user representation)(hierarchical recurrent network) User Representation
- 24. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Stateof theArt Performance 0.954 0.928 0.925 0.922 0.91 0.856 Rolling Average T-SVD [2009] PMF [2008] RRN [2017] DeepRec [2017] HRNN Ratings RMSE on Netflix 98 MM interactions, 500k users, 18k items Rolling Average T-SVD [2009] PMF [2008] RRN [2017] DeepRec [2017] HRNN 0.933 0.916 0.871 0.857 0.846 Rolling Average FM [2012] I-AutoRec [2015] RNN HRNN Ratings RMSE on MovieLens 20 MM interactions, 173k users, 131k items Rolling Average FM [2012] I-AutoRec [2015] RNN
- 25. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Large ParameterSpace ‒ HPO andAutoML SIMS DeepFM HRNN Automatic within algorithm parameter tuning (HPO) Automatic algorithm selection (AutoML) (time decay) (depth, size) (depth, size, height)
- 26. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. WhyAmazon Personalize? (hierarchical sequential models) work well for many personalization tasks (stable) solutions out of the box Includes many tricks from deep learning practitioners AutoML and HPO Best possible results (on your data) with minimal intervention Bring your own recommender Tune Amazon Personalize to (your specific needs)
- 27. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Personalizing Customer Engagement at Scale Mallika Krishnamurthy Global Head, Strategy & Insights Domino’s Pizza Enterprises
- 28. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 29. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Breakout repeats Day of Week, Month Day Session Title Time – Time | Location Day of Week, Month Day Session Title Time – Time | Location Day of Week, Month Day Session Title Time – Time | Location Australia | New Zealand | Japan | Germany | Netherlands Belgium | Luxembourg | France 2400+ stores
- 30. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Largest Pizza Chain in Australia
- 31. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Clicktoadd slidetitle(size48) L E A D I N G T H E I N T E R N E T O F F O O D I N E V E RY N E I G H B O R H O O D
- 32. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Technology focused
- 33. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AI and machine learning A C C E S S I B L E A N Y W H E R E
- 34. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Owned media Channels
- 35. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 36. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Online ordering at scale
- 37. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 38. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 39. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Personalization of deals
- 40. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 41. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 42. Thank you! © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Mallika Krishnamurthy Global Head, Strategy & Insights Domino’s Pizza Enterprises
- 43. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Mike Pyland CTO RBmedia / Recorded Books Inc. rbmediaglobal.com
- 44. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Global leader in spoken audio content and digital media distribution technology that reaches millions of consumers — at home, in the car, and wherever their mobile devices take them. RBmedia produces exclusive titles and delivers the finest digital content and information — including audiobooks, eBooks, educational courses, entertainment titles, and much more.
- 45. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Library Digital Media Delivery Platform 5,000+ public and academic libraries worldwide Follows standard Library Checkout/Hold model Checkout growth 50+% over last year >700K Titles Audiobooks eBooks Magazines & Comics Streaming Video Education Health & Wellness
- 46. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Current Recommendations Simple presentation of most popular titles Default is Romance and Mystery Based on genre User must select genres Removes previously checked out books
- 47. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Current Recommendations Simple presentation of most popular titles Default is Romance and Mystery Based on Genre User must select genres Removes previously checked out books Challenges No personalization Everyone gets same titles worldwide Just Audiobooks/eBooks Not sure there is enough benefit to customer Don’t want to add clutter
- 48. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. NewRecommendations Engine Goals Quick Time to Market Didn’t want to analyze models for 6-12 months
- 49. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. NewRecommendations Engine Goals Quick Time to Market Didn’t want to analyze models for 6-12 months Low Cost & Scalable No Data Scientist needed API Integration Delta updates Vs Full Loads
- 50. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. NewRecommendations Engine Goals Quick Time to Market Didn’t want to analyze models for 6-12 months Low Cost & Scalable No Data Scientist needed API Integration Delta updates Vs Full Loads Personalized Similar users & similar books Real Time (New Users handled) Preferences – eAudio or eBook, Available Flexible Future Ability to adjust results & models – Change weight slightly Add additional media types and data
- 51. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Personalize Goals Achieved Quick Time to Market In <1 month we had prepared data, loaded datasets, trained model, saw results Low Cost No Data Scientist – Easy to follow steps so any developer can deliver API Integration for real time Delta updates Only pay for GPUs when in use Don’t need to do 5000 models for different libraries
- 52. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Personalize Goals Achieved Quick Time to Market In <1 month we had prepared data, loaded datasets, trained model, saw results Low Cost No Data Scientist – Easy to follow steps so any developer can deliver API Integration for real time Delta updates Only pay for GPUs when in use Don’t need to do 5000 models for different libraries Scalable Auto sizing for processing Ability to extract similar data and load into other data repositories like search Easily leverage AWS services (Lambda, Kinesis, and API Gateway) Real Time / Delta loads
- 53. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Personalize Goals Achieved Personalized Similar books, Similar users, results based on usage data Preferences learned – eAudio or eBook, Genre Real Time (New Users) with minimal inputs
- 54. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Personalize Goals Achieved Personalized Similar books, Similar users, results based on usage data Preferences learned – eAudio or eBook, Genre Real Time (New Users) with minimal inputs Flexible Future Ability to easily choose recipes/models to test best results Full SageMaker and AI capabilities, if desired Ability to expand Content metadata and relationships Ability to add sentiment data Ability to mix/filter with other information in API results Ownership & Availability Full control
- 55. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Personalize ‒ Process Choose an MVP What problem are you trying to solve??? Choose data for Customer, Products, Relationship(Checkout/Sales) Review data – cleaning up data may be the hardest part Start Small (doesn’t have to be all of your data to start) Define formula for “strength” of relationship – AI may fix initial thoughts Walk through the Notebook for understanding of steps Choose the top recipes for desired results (or try them all and review) Prepare and Load data to S3 Run the Notebook Load Datasets Define “Solution” & Train – run multiple recipes to find best one Test outputs Operationalize – Full data, real time updates, APIs for final use case
- 56. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Personalize Delivered Results In <1 month we knew it would deliver expected results = Win for IT Gives our users truly personalized results = Happiness Gives our libraries comparable / competitive tool = Engagement Competitive advantage for RBdigital = ROI Thank you to the AWS team for delivering a much needed tool
- 57. Thank you! © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Mike Pyland mpyland@recordedbooks.com www.linkedin.com/in/mikepyland
- 58. Learn more and signup for the Amazon PersonalizePreview: https://aws.amazon.com/personalize/
- 59. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Hinweis der Redaktion
- Current systems only employ factorization machines or item similarity models which do not respond to user intention in real time leading to recommendations which lead to poor user experience Source for 1: https://drive.corp.amazon.com/documents/AMLC2018/AcceptedPapers/amlc_recurrent_bandit.pdf - Rule based systems involve segmenting customers according to their demographic information (such as age) and category interests. These manually curated segments and personalization actions soon balloon to thousands and it becomes hard to validate performance of new rules as A/B tests take time to converge. In Amazon our ML systems ALWAYS beat human curated rules Source for rule based: https://drive.corp.amazon.com/documents/AMLC2018/AcceptedPapers/amlc2018-bjd.pdf
- Headquartered in Australia, Domino's Pizza Enterprises Ltd (DPE) is the largest franchisee for the Domino’s Pizza brand worldwide.
- DPE is located in 8 markets including Australia, NZ, Belgium, France, the Netherlands, Japan, Germany and Luxembourg, with more than 2400 stores
- It is the largest pizza chain in Australia in terms of both network store numbers and network sales. And our market share in pizza is dominant.
- Our Vision is to Lead the Internet of Food in Every Neighborhood.
- To achieve this, there has been an ongoing commitment to being at the forefront of technological innovation that focuses on making the customer experience a better one. It led to the development of GPS Driver Tracker that launched in 2015, that lets a customer track their pizza from the store to their door. - More recently we launched drone deliveries - our autonomous delivery vehicle, to add to our fleet of delivery methods. And our New Pizza Checker with Augmented Reality – allows users to visually make their pizza by dragging toppings onto a virtual pizza. Creates a more engaging and true to life ordering experience for our customers.
- How we communicate with our customers through technological advancements in AI and Machine Learning is an area we have been actively exploring.
- We utilize our owned-media channels to reach our customers on an ongoing basis and this has been highly effective for us because our customer base is so engaged.
- And while we’ll continue to focus on growing our subscriber base, we have a major opportunity to truly personalize the engagement we have with our customers, at scale.
- As over 60% of our orders are placed online, we have built up rich customer data sets that enable us to deeply understand who are our customers and how they behave
- Today, we are sending our customers SMSs, eDMs and push notifications, but our targeting is too manual. We are missing the automation to scale our testing framework and the machine learning to provide the intelligence behind the targeting. Amazon Personalize integrated with Pinpoint has enabled us to start testing at a rapid pace.
- Time based test: Time between when the message was sent and the message was clicked was on average 55 minutes, with the test group, but within the control group the average was 88 minutes. Meaning the personalized SMS was more timely and conversion was marginally better.
- We’ve also seen an uplift in conversion by approx. 0.07% percentage points when sending personalized deals to our test group versus standard deals to our control group.
- Time was our main challenge. We kicked off the POC 3 weeks before the Conference, spent 1 week developing and getting it up and running, and two weeks testing and improving…so there’s a lot more opportunity for us to test further and fine tune our model.
- On the flip side, it’s fairly impressive that we were able to do so much so quickly. A testament to AWS’s platforms and the teams working behind the scenes both in Seattle and in Brisbane. So Why go on this journey with AWS and Amazon Personalize? Well we wanted to work with a partner who aligns with our values. A partner that would embrace our fast fail mentality, working with us to achieve outcomes at a rapid pace; A partner who is not afraid to imagine the possibilities with us And most importantly, a partner who puts the customer at the heart of everything they do.
- RB is known for high quality Audiobooks since 1979 Audiobooks.com is #2 in the market next to Audible– we have a great relationship with Audible I believe Rbmedia’s 35+K audiobooks makes us the largest publisher on Audible WFH and Wavesound cover our UK and Australia markets for exclusive audiobooks and library distribution Tantor, CA, Gildan and Highbridge produce over 4000 titles a year – data driven Gildan and our exclusive partners give us more business audio titles than any other publisher
- RBdigital is a multi-tenant hosted platform for public libraries Follows library model for Checkouts and holds Each library owns their selected content May share content from a System or Consortia Recommendations must follow the list of owned/shared content and should be Available for checkout now The RB All You Can Eat model helps with availability by providing more than 25K titles as Unlimited access RBdigital launched in 2012 with just Audiobooks, adding eBooks in 2015 then Magazines, Comics, Educational resources and SVOD
- Like most home grown personalization systems we did a basic most popular list Based it on Genre to show category patron likes Filters previously read
- It isn’t personalized to the individual Have to ask patron or they automatically get Romance/Mystery Last board dinner meeting I was told “it better not bring up Romance or it’s worthless” Buried due to lack of confidence it helps Checked the box for marketing, but did it truly help the user?
- I needed to sell executive team on value prior to investment Prove Value – small enough to just steal dev resources for PoC / MVP
- No upfront commitment to provider for $100K investment or multiple years Had to work with our services Can’t add more batch jobs at night – real time
- Had to be better than today – more personalized Work with limited data for new users Needed to be able to add our own filtering to remove UnOwned, UnAvailable, and previously read books Had to be able to adjust for the future content (mag/video), user behaviors (detail clicks, searches…) Need control - to be able to Boost for seasonal (Christmas, movie release, …) So we turned to AWS for a solution
- In 3 weeks we had results. Now we could do it again in less than 1 week. 1 DBA to get data, 1 Dev part time to test It was enough to know we had value and get moved into sprint cycle No Data Scientist needed, sorry buy you guys can analyze in circles for months… If needed, Data Scientist investment based on $ – not locked into a black box Easy/Cheap to maintain - self learning – no humans involved in daily data loads/training Pre-built tools - JavaScript client and server-side Java/.Net tools Only Pay based on use – lower cost than running your own Shown Consolidated model reduces complexity and increases accuracy
- As with all AWS services it Scales to future load Fast learning – real time data load Can still leverage cached results based on API layers for speed and cost
- Noticeably better than what we had before (5X better) This means we are going from 2 good books to 10 books out of 25? It seemed to “learn” things like Genre even before we added the data – AI works
- AWS has packaged a great set of out of the box solutions to start, but the flexibility is wide open Now I see value of Sagemaker – too many “terms” flowing last year for me to see as manageable Practical implementation brings it home that we can build and invest Dev or Data Scientist can improve/tweak what was completed quickly over long run Add additional intelligence over time Ability to add real time additional users, relationships/interactions, and content We have control over how we use and display results Use in new ways – anything curated/personalized or with similar patterns will work Really nice building block for your environment
- Keep it Simple / high level – Data Science/AI sounds expensive so tout it after it works - impressive Do a PoC to show value and convince executive team to invest Imported “clean” data to start ~1GB (unique values, don’t need text – Just IDs not names/email – no PII) Pick 3+ users(self and coworkers) and get back top 10 items – more than stats – results people can see Review precision across multiple recipes – quick and simple to test out different models If it doesn’t look right review recipe and data – does it make sense – optimize Don’t give up based on 1 result – test more users and recipes Get to Market and Operationalize then Start adding things like A/B testing, Clicked On – Implicit Vs Explicit? Process not 1 time solution Make the baby come alive – interaction with results will train model faster
- Happy IT Happy Librarians Happy Patrons Happy Execs