3. Common applications& use cases
Personalized
recommendations
Search
reranking
Notifications and
emailsRelated Items
4. 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
5. 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
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. DeepLearningdeliversstateof theartperformance
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
19. Traditionaltime-seriesmodels
• Independent forecasts
• Strong structural assumptions
• De-facto industry standard
• Well-understood, > 50 yrs. research
• Data must match the structural
assumptions
• Cannot identify patterns
across time series
21. Traditional methods struggle with real-world forecasting
Can’t handle
time-series with
no history
Only process a
single time-
series at a time
Don’t consider
additional inputs:
related time-series,
metadata
Only predict a single
value: how trustworthy
is it?
26. Usingadditionalinputs
• Additional inputs can
• Explain historical data
• Drive forecast behavior
• Examples from retail
• Price information
• Information about promotions
• Out-of-stock information
• Web page views
• Known future events
• Categorical inputs can be used to
identify group-level patterns
Fashion
Women’s
Clothing
Shoes
Watches
Men’s
Clothing
Shoes
Watches
Girls'
Clothing
Shoes
Watches
Boys'
Clothing
Shoes
Watches
30. Probabilisticforecasts
• Quantification of uncertainty
• Support optimal decision making
• Make “wrong” forecasts useful
• Forecasts can be obtained for
different quantiles of the predictive
distribution
p10: 10% of predictions with be lower
p50: the mean value
p90: 90% of predictions with be lower
p10-p90 interval: 80% of possible predictions.
31. Deeplearningtime-seriesmodels
• Global models: identify patterns using
all available time series
• Group-dependent seasonality and lifecycle
• Behavior in response to extra inputs
• Weak structural assumptions
• Can be significantly more accurate
than traditional methods
• Can easily incorporate and learn from
rich metadata
• Support cold-start forecasts for new
items
This talk is for builders
I want to build a new app or website and I want it to work on every platform
I want easily leverage AWS from my existing web or mobile apps and I don’t want to rewrite everything
I want to learn about cool new development tools like React, GraphQL, CLIs, and serverless technologies
I want to focus less on ops and configuration and more on my product
If you recognize yourself in one of the previous sentences, you’re in the right room.
TODO for the presenter : adjust the banner to the conference you will speak to.
LAUNCH CUSTOMERS: Domino’s, Navitime, Rbmedia/Recorded Books Inc., Spuul, Zola.
One of our launch customers, Domino’s Pizza, is using Amazon Personalize to predict purchasing behavior and apply context about individual customers and their circumstances to deliver personalized promotions and notifications.
Spuul, a video streaming platform delivering Indian movies and TV shows to an audience worldwide, was manually categorizing and displaying content to users based on broad segmentation. With Amazon Personalize, they now can provide unique and personalized recommendations to each customer.
And Sony Interactive Entertainment (SIE), which is Sony’s video game division overseeing the PlayStation ecosystem, is using Amazon SageMaker and Amazon Personalize to automate and accelerate their machine learning development, and drive more effective personalization at scale.
Another launch customer, Zola, develops innovative wedding planning tools to serve couples. They want to provide the best possible recommendations to our customers based on their style, interests, or preferences. Until now, those recommendations were implemented via rule-based ranking, popularity, or, more recently, via a similarity model calculated offline. With Amazon Personalize, Zola can respond to customer actions in real-time and quickly deliver solutions that would have otherwise taken a much larger team and several months development time.
NAVITIME, a leading provider of navigation technology and services in Japan, is using Amazon Personalize to to improve the accuracy of predictions in their navigation app by personalizing search results as well as recommended navigation routes, based on the users personal preference.
1/ Personalize works very similar to Forecast. You provide data, which in this case is activity stream data from a web application or mobile app, information about the available inventory of things to recommend, and user demographics if known
2/ Personalize loads the data, inspects it, identifies features, selects algorithms and hyperparameters, and then trains and optimizes models
3/ Because Personalize is able to take full advantage of all of the available data to train, the models can be very large, in the gigabytes
4/ We put to use a lot of the lessons we’ve learned over the years to operationalize these sophisticated models to implement a separate feature store for the model data and automatically scale the models themselves to keep latency below 100ms
5/ All of this happens automatically for the user
6/ At the end of the process, users have a customized personalization API that’s been trained on their unique data
LAUNCH CUSTOMERS: Mercado Libre, CJ Logistics
Customers like MercadoLibre, Latin America's most popular e-commerce site. is using Amazon Forecast to predict demand for over 50,000 different products, Forecast’s state-of-the-art deep learning algorithms available out of the box. Forecast is removing all the heavy lifting of setting up pipelines, re-training schedules, and re-generating forecasts, so they can experiment with hundreds of models very easily.
1/ To use Amazon Forecast, you provide access to historical data, plus any data which you believe may be useful in making predictions
2/ Amazon Forecast loads the data, inspects it, identifies features, selects algorithms and hyperparameters, and then trains and optimizes models before hosting them in a high availability environment
3/ All of this happens automatically for the user
4/ At the end of the process, users have a customized forecasting API that’s been trained on their unique data
5/ Forecast can forecast any time series scenario, including retail demand, travel demand, AWS usage, revenue, web traffic, and advertising demand
TODO for the presenter : adjust the banner to the conference you will speak to.