Upload
sri-ambati
View
650
Download
0
Embed Size (px)
Citation preview
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Self Guiding ApplicationsVenkatesh Yadav
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Why ?
If an app takes more than a few seconds to learn, a majority of users are going to
uninstall (MOBILE).
Creating that engaging initial use experience is challenging, predominantly
constrained by
Complexity of the app
Screen space
Etc.,
2
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Why ?
We want to provide best “Engagement Experience” possible using crowd sourced
application usage data.
3
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Self Guiding Applications
4
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential. 5
Self Guiding Applications95
%
Action 1
92
%
Action 2
88
%
Action 3
85
%
Action 4
80
%
Action 5
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Phase 1 – Proof Of Concept
6
The Setting
A mobile photo editing app.
Relatively less complicated – approx. 20 possible actions
Constrained in space – ribbon scroll and searching for actions
The Goal
Create engaging user experience, minimize scrolling and searching
Predictive Feature Panel and Contextual Window
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Crowdsourced Product Usage Data
7
Each row is a set of actions (like a workflow) performed in an image editing session
Total 100K rows of data, of approx. 20 possible actions
001 002 003
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Machine Learning Algorithms
8
Markov Model
Next action is dependent only on the present action, not on the sequence of previous actions
N Action Bayesian Model
Next action is dependent on the sequence of past N actions
N Action Neural Network Model
Next action is dependent on the sequence of past N actions
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Predictive APIs
10
Evaluated few approaches to expose predictive models via HTTP REST APIs
Loose coupling between model creation and consumption
Continuous model development and deployment capability
Create Java POJO for the predictive
model
Wrap POJO by any Java Web API
framework
Can better integrate with Adobe.io
platform
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
REST API – H2O
11
Training
Dataset
Platfor
m
Server
Train Model
Data Science
Team
Smart Apps
Predictiv
e Model
(Java)
Predictive
API (Jar
file)
© 2015 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Demo
“Once models are deployed to the platform, they can
begin receiving API requests and sending predictions back
to the applications.”
Loose coupling
Continues model development
Provide REST API interface to predictive model
(Demo)
Integration into an application (Next Step)