Content Complexity, Similarity, and
Consistency in Social Media:
A Deep Learning Approach
Gene Moo Lee
University of Texas at Arlington
Joint work with
Donghyuk Shin (UT Austin/Amazon), Shu He (UConn),
Andrew B. Whinston (UT Austin)
DSI 2016, Austin TX
Social media: More users
2
Social media: More spending
3
Challenges and opportunities: 78% photos
4
Source: Chang et al. 2014
Research questions
• How can firms optimize social media strategies by
incorporating visual content?
• Specifically, what are the determinants of consumer
engagement in terms of “likes” and “reblogs” (sharing)
actions?
• How visual and textual contents play role?
• Operationally, how to construct measures on these
unstructured data sources?
5
Tumblr data
• Tumblr: microblogging platform (acquired by Yahoo!)
• 35,651 posts by 183 companies (May - Oct 2014)
• Automobile, Entertainment, Food, Fashion,
Finance, Leisure, Retail, Tech
• 89.7% photo & text, 6.3% pure text, 4% videos
• Collected “likes” and “reblogs” until Apr 2015
6
Company blogs in Tumblr
7
BMW USA Vogue IBM
Data: blog post and engagement
8
Post = Visual Info (Image) + Textual Info (Text, Tags)
Customer engagement = Notes (Likes + Reblogs)
Visual features
• Aesthetics (beautiful photos)
• Adult-contents
• Celebrity
• Feature complexity (low-level, flashy images)
• Semantic complexity (high-level, complex meaning)
• Number of salient objects
9
Feature complexity (low level)
• Visual complexity theory [Donderi 2006a, Pieters et al. 2010]
• Visually complex (flashy) images (colors, luminance,
shape) gets more attention
• This feature complexity can be captured by the
image’s compressed file size [Donderi 2006a; Donderi
2006b; Machado et al. 2015; Forsythe et al. 2011]
• However, this complexity can only capture low-level
complexity based on “pixel” values
10
Semantic complexity (high level)
• Recognition-By-Components theory [Biederman 1987]
• Human object recognition is invariant to feature
factors (colors, brightness, edges, positions, etc.)
• Vessel and Rubin (2010) show that visual preferences
are influenced by semantic content in the image
• We posit that semantic complexity matters!
• Operational question: How do we calculate semantics
from unstructured images?
11
Deep learning
• A branch of machine learning, inspired by human brain
• Algorithms to model high-level abstractions with multiple processing
layers of non-linear transformations
• (1) theoretical breakthroughs, (2) Big Data, (3) powerful computation
• Successfully applied in image/video/voice recognition, AlphaGo, etc.
12
Semantic complexity via deep learning
• Deep convolutional neural network (CNN) [Jia et al. 2014]
• Model trained with 1.2 million images with tags (ImageNet, Flickr)
• Tested on 53,417 images from brand-generated Tumblr posts
• Each image is represented by a 1,700 dimensional vector, where each
value is the confidence score w.r.t. an object (tag)
• We define semantic complexity as the Shannon Diversity Index (entropy)
on the 1,700-dimensional vector
• max = log(d), if p is uniformly distributed
• min = 0, if p_i = 1 for some i
13
ImageNet: Image DB with tree-structure tags
14
Source: ImageNet
More visual features
• 7th-layer output = robust representation of the image for “computer vision” tasks
• Aesthetic/beauty score [Dhar et al. 2011 (CVPR, Vision)]
• Adult-content score [Sengamedu et al. 2011 (MM, Vision)]
• Celebrity (450 celebrities) [Parhki et al. 2015 (BMV, Vision)]
• Number of salient objects [Zhang et al. 2015 (CVPR, Vision)]
15
Examples: Visual features
• Visual complexity theory (Attneave 1994,
Donderi 2006, Pieters et al. 2010)
• Visual stimuli are a composite of
colors,luminance, shape, number of
objects/patterns
16
Textual features
• Two textual sources: text and tags
• Length: # of words, # of tags
• Topic complexity: LDA topic model (text, tags)
• Order complexity: word2vec (for text only)
17
Examples: Textual features
• Topics
• Word clusters
18
Visual-Textual Content Similarity
• Image: pixels, Text/Tags: characters
— Need a common representation!
1. Represent each image as a collection of the predicted labels
obtained from deep learning — “image corpus”
2. Train LDA with both image and text/tags corpora — topic
distribution for images and text/tags
3. Cosine similarity between the two corresponding topic
distribution
19
Examples: Content similarity
• Topics
• Word clusters
20
21
Empirical Model
• Linear fixed effects model
• DV (likes/reblogs): take log transformation due to their
skewed distributions
• Capture blog (firm) heterogeneity
• Capture time effects (day of week, month)
• Other models
• Identical results with random effects
• Consistent results with negative binomial model
22
23
Summary and implications
1. Large-scale analysis on visual content in social
media
2. New visual semantic complexity via deep learning
• Able to relate visual and textual content
Visual content analysis can be used to optimize
content design for social media marketing
24
Thank you!
Contact Info: Gene Moo Lee
gene.lee@uta.edu

Content Complexity, Similarity, and Consistency in Social Media: A Deep Learning Approach

  • 1.
    Content Complexity, Similarity,and Consistency in Social Media: A Deep Learning Approach Gene Moo Lee University of Texas at Arlington Joint work with Donghyuk Shin (UT Austin/Amazon), Shu He (UConn), Andrew B. Whinston (UT Austin) DSI 2016, Austin TX
  • 2.
  • 3.
  • 4.
    Challenges and opportunities:78% photos 4 Source: Chang et al. 2014
  • 5.
    Research questions • Howcan firms optimize social media strategies by incorporating visual content? • Specifically, what are the determinants of consumer engagement in terms of “likes” and “reblogs” (sharing) actions? • How visual and textual contents play role? • Operationally, how to construct measures on these unstructured data sources? 5
  • 6.
    Tumblr data • Tumblr:microblogging platform (acquired by Yahoo!) • 35,651 posts by 183 companies (May - Oct 2014) • Automobile, Entertainment, Food, Fashion, Finance, Leisure, Retail, Tech • 89.7% photo & text, 6.3% pure text, 4% videos • Collected “likes” and “reblogs” until Apr 2015 6
  • 7.
    Company blogs inTumblr 7 BMW USA Vogue IBM
  • 8.
    Data: blog postand engagement 8 Post = Visual Info (Image) + Textual Info (Text, Tags) Customer engagement = Notes (Likes + Reblogs)
  • 9.
    Visual features • Aesthetics(beautiful photos) • Adult-contents • Celebrity • Feature complexity (low-level, flashy images) • Semantic complexity (high-level, complex meaning) • Number of salient objects 9
  • 10.
    Feature complexity (lowlevel) • Visual complexity theory [Donderi 2006a, Pieters et al. 2010] • Visually complex (flashy) images (colors, luminance, shape) gets more attention • This feature complexity can be captured by the image’s compressed file size [Donderi 2006a; Donderi 2006b; Machado et al. 2015; Forsythe et al. 2011] • However, this complexity can only capture low-level complexity based on “pixel” values 10
  • 11.
    Semantic complexity (highlevel) • Recognition-By-Components theory [Biederman 1987] • Human object recognition is invariant to feature factors (colors, brightness, edges, positions, etc.) • Vessel and Rubin (2010) show that visual preferences are influenced by semantic content in the image • We posit that semantic complexity matters! • Operational question: How do we calculate semantics from unstructured images? 11
  • 12.
    Deep learning • Abranch of machine learning, inspired by human brain • Algorithms to model high-level abstractions with multiple processing layers of non-linear transformations • (1) theoretical breakthroughs, (2) Big Data, (3) powerful computation • Successfully applied in image/video/voice recognition, AlphaGo, etc. 12
  • 13.
    Semantic complexity viadeep learning • Deep convolutional neural network (CNN) [Jia et al. 2014] • Model trained with 1.2 million images with tags (ImageNet, Flickr) • Tested on 53,417 images from brand-generated Tumblr posts • Each image is represented by a 1,700 dimensional vector, where each value is the confidence score w.r.t. an object (tag) • We define semantic complexity as the Shannon Diversity Index (entropy) on the 1,700-dimensional vector • max = log(d), if p is uniformly distributed • min = 0, if p_i = 1 for some i 13
  • 14.
    ImageNet: Image DBwith tree-structure tags 14 Source: ImageNet
  • 15.
    More visual features •7th-layer output = robust representation of the image for “computer vision” tasks • Aesthetic/beauty score [Dhar et al. 2011 (CVPR, Vision)] • Adult-content score [Sengamedu et al. 2011 (MM, Vision)] • Celebrity (450 celebrities) [Parhki et al. 2015 (BMV, Vision)] • Number of salient objects [Zhang et al. 2015 (CVPR, Vision)] 15
  • 16.
    Examples: Visual features •Visual complexity theory (Attneave 1994, Donderi 2006, Pieters et al. 2010) • Visual stimuli are a composite of colors,luminance, shape, number of objects/patterns 16
  • 17.
    Textual features • Twotextual sources: text and tags • Length: # of words, # of tags • Topic complexity: LDA topic model (text, tags) • Order complexity: word2vec (for text only) 17
  • 18.
    Examples: Textual features •Topics • Word clusters 18
  • 19.
    Visual-Textual Content Similarity •Image: pixels, Text/Tags: characters — Need a common representation! 1. Represent each image as a collection of the predicted labels obtained from deep learning — “image corpus” 2. Train LDA with both image and text/tags corpora — topic distribution for images and text/tags 3. Cosine similarity between the two corresponding topic distribution 19
  • 20.
    Examples: Content similarity •Topics • Word clusters 20
  • 21.
  • 22.
    Empirical Model • Linearfixed effects model • DV (likes/reblogs): take log transformation due to their skewed distributions • Capture blog (firm) heterogeneity • Capture time effects (day of week, month) • Other models • Identical results with random effects • Consistent results with negative binomial model 22
  • 23.
  • 24.
    Summary and implications 1.Large-scale analysis on visual content in social media 2. New visual semantic complexity via deep learning • Able to relate visual and textual content Visual content analysis can be used to optimize content design for social media marketing 24
  • 25.
    Thank you! Contact Info:Gene Moo Lee gene.lee@uta.edu

Editor's Notes

  • #24 Industry subsample analysis Long- and short-term customer engagement Categorize posts/blogs into ‘utilitarian’ vs ‘hedonic’ Examine non-linear effects