Chief Scientist at Tooploox
Assistant Professor at Warsaw University of Technology
Online content
popularity prediction
Tomasz Trzciński
Bio
PhD in Computer Vision
Internships at Google,
Qualcomm & Telefonica
VP R&D, PlayfulVision
Guitar player & proud father
http://pennystocks.la/internet-in-real-time/
i
How many will we get?
~22 likes
per day*
* http://popularity.csail.mit.edu/
What makes an image popular?
2.3M images from Flickr with
#views grouped by users
[What makes an image popular? A. Khosla, A. Das Sarma and R.Hamid, WWW’14]
Color?
[What makes an image popular? A. Khosla, A. Das Sarma and R.Hamid, WWW’14]
Visual cues?
edge/gradient
histograms
convolutional neural
network responses
bag-of-words
representation
[What makes an image popular? A. Khosla, A. Das Sarma and R.Hamid, WWW’14]
Social network?
#
#
#
#
#
#
mean views
of the user
images
uploaded
group
memberships
contacts
tags
title length
[What makes an image popular? A. Khosla, A. Das Sarma and R.Hamid, WWW’14]
Results
demo: http://popularity.csail.mit.edu/
[What makes an image popular? A. Khosla, A. Das Sarma and R.Hamid, WWW’14]
Results
[What makes an image popular? A. Khosla, A. Das Sarma and R.Hamid, WWW’14]
Selfies
How to take a good great #selfie?
[What a deep NN thinks about your #selfie? A. Karpathy, Blog.
Let’s ask a deep convolutional neural network!
How to take a good great #selfie?
[What a deep NN thinks about your #selfie? A. Karpathy, Blog.
Results: Best selfies
[What a deep NN thinks about your #selfie? A. Karpathy, Blog.
Top-100 Top Male
Results: Worst
[What a deep NN thinks about your #selfie? A. Karpathy, Blog.
Framing your #selfie
[What a deep NN thinks about your #selfie? A. Karpathy, Blog.
Sometimes a good selfie… is a selfie without you at all
Let’s record a viral movie…
220 million views
Early patterns reflect long-term interest
[Predicting the popularity of online content. G. Szabo and B. A. Huberman, ACM’10]
Prediction ~ Regression
Univariate Linear (UL)
Multivariate Linear (ML)
ML + Radial Basis Function
Support Vector Regression
[Using early patterns to predict the popularity of YouTube videos. H. Pinto et al., WSDM’13]
[Predicting popularity of online videos using Support Vector Regression. T. Trzcinski and P. Rokita, sub. TCSVT’15]
Φ(x, y) = exp
✓
−
||x − y||2
2σ2
◆
∀
t<T
viewsUL(v, T) ∼ ln (views(v, t))
∀
t1,...,tn<T
viewsML(v, T) ∼ ln (views(v, t1), ..., views(v, tn))
viewsSV R(v, T) ∼ Φ (views(v), ..., views(SV ))
viewsRBF (v, T) ∼ viewsML + Φ (views(v), ..., views(random))
Prediction before publication
Visual cues only
Video length
Dominant color
Scene dynamics
Text (OCR)
Faces
Clutter
Thumbnails
[Predicting popularity of online videos using Support Vector Regression. T. Trzcinski and P. Rokita, sub. TCSVT’15]
Opening Scene
Top-20
Worst-20
Results
[Predicting popularity of online videos using Support Vector Regression. T. Trzcinski and P. Rokita, sub. TCSVT’15]
& Data Science
Data
Science
viewers’ interests
virality prediction
competition analysis
trends across platforms
demographics
reverse-engineer
WE’RE HIRING!
Data Scientists
Machine Learning, Predictive Analytics, Data Visualization, Social Networks
THANK YOU!
www.tooploox.com/jobs

Online content popularity prediction