Consumers spend an enormous amount of time on their social media channels. That’s why it is now the major focus of many marketing departments. So whether your marketing department tweets, snaps, shares, pins, or grams or all of the above your consumers are ever increasing there engagement with you through these channels. Social media is a visual medium. So the image selection process should be a pretty sophisticated, right? How engaging are the images you use in your social marketing? Do you know which images or types of images will get the most engagement? I will show you how you can leverage the power of Artificial Intelligence (AI) to understand what types of images on your social channels can drive the most engagement. In this session, I will demo AI Image classification process and show how these classes can be used to predict engagement.

1. How Engaged Are your
Posts?
Social Image Analytics
Jack McCush – Teradata
#SAISAI9

2. Problem Statement:
• Marketing has reactive tools to optimize revenue available
through social media
• Enable predictive consumer engagement across image and
account factors
Theoretic Customer Scenario:
• Priceline ($12.7 B revenue) looking to catch Expedia
• 30% of $4.3 B current digital spend on social media
• ROI Discussion: 2-4% sales growth, or ~$380 M above
current growth trends with image optimization
Overview
#SAISAI9

3. Introducing Project Cyclops
Model
Training
Model
Scoring
Apply
ResNet50
Apply
ResNet50
Predicted: [
('seashore', 0.8395325),
('lakeside', 0.14620699),
('breakwater', 0.0025936835),
('pier', 0.0018090468),
('picket_fence', 0.0015644263)]
• Top 5 Classes for each
image
• Class is attribute,
probability of class is
attribute value
• With 50K+ images, 997
classes represented
Train
XGBoost
Score
XGBoost
XGBoost_score.py
XGB.pickle
XGBoost_score.py
XGB.pickle
= 2,300
Predicted Likes
#SAISAI9

4. Provide the end user with a REST API where
they can receive a prediction of a given image
impact:
1. Download images from accounts into an
AWS S3 storage.
2. Deploy and train the model using
AnalyticOps.
3. Dockerize the trained model.
4. The REST API invokes the prediction
and returns the insight.
Architecture Deployment Methodology
#SAISAI9

5. Demo!
#SAISAI9

6. Continued Development: Project Cyclops
Oil & Gas:
Structural Inspection
Healthcare:
Dermatologic Severity
Security:
Smart Camera; Smart City
Retail:
B to B to C
#SAISAI9