Intelligence driven automation of user communications is aimed at determining what kind of communications should be sent to whom, when, and how. In addition to better user engagement and cost analysis, this framework unlocks several critical use cases for Uber. These include user targeting for marketing and other campaigns, appropriate content and channel routing, offensive content discovery, delaying or blocking marketing communications in events of disasters, relevance sorting of communications based on user interests, etc.
We will talk about how we are using data science analysis to understand our communications data and build classification models to categorize our content at scale and using those predictions for segmentation, automation, and tuning communication delivery. We are also exploring deep learning techniques to achieve this objective. This talk will include some of the architecture details around model training, deployment and predictions as well as longer term plans on how we envision this platform.
3. 01 Mission
02 Communication Terminology
03 Vision for Intelligent Communications
04 Architecture
05 Scale
06 Use Cases
07 Q&A
4. Drive billions of individual interactions intelligently and efficiently,
across all communication channels that customers use, and
adapt in real time to their behaviors.
WHY | MISSION
8. Suggested channels:
● Based on metadata
● Engagement metrics
● Cost ROI
● ...
Campaign planning:
● Previous campaigns
● Engagements
● Experimentation
● Schedules
Personalization:
● Channels/Medium
● Formatting
● ...
In-flight rerouting:
● Channel Affinity
● Real-time feedback
● Marketplaces
● ...
Vision: AI-driven Marketing and CRM campaigns
High level objective:
(human)
● Type of campaign
● Objective function
● Target audience
● Budget
9. Tools
API
Clients
Gateway
Comms Data Flow
Intelligence
Recommendations: Channel, Content,
Target, Schedule
Suppression: Unsubscribes, Preferences
Comms Providers
Delivery
Channel specific
delivery
Post Delivery
Delivery Metrics,
Data Aggregation
Pre Delivery
Author, Schedule, Target
DataLayer
Feature store
Content stores
Raw Data
storage
10. 3.5BMessages Sent Daily
500M Events Processed Daily
Communication Scale
● 500M messages sent and
3.5B message related
events processed through
the system
● Transactional vs. Marketing
SLAs
● Global: Supports 168 locales
Scale
11. Suppression
Framework
City Supp.
User Supp.
Hard Bounce
Control Groups
Freq. Rate Limit
Personalization
…
Filters: Suppression Framework
Blocking communications in-flight ...
SYNC: should send?
Gateway
holdout list
… to
Providers
Batch
processing
Uber services
Uber services
Users C* Experiments
ASYNC handles async actions
12. ● Estimate most likely city for a
user
● Disperse features to Cassandra
via Spark app periodically
● Suppressor framework stops
marketing messages in
real-time
Example: City-based message suppression
Scheduled
Spark JobHive
Cassandra
Suppressor
shouldSend()
User
Features
Everyday
Feature Dispersal