Download presentation
Presentation is loading. Please wait.
1
Personalized Social Image Recommendation
Jianping Fan Department of Computer Science University of North Carolina at Charlotte
2
Social Image Sharing, Tagging, Friendship, ….
3
Social Image Sharing, Tagging, Friendship, ….
4
What people may care? Entertainment Tagging Interesting Photos
Making friends by tagging & sharing Knowing hot topics Navigating relevant photos ……
5
Significant Difference with News
Social Images & Tags Information are provided by non-professionals Information filtering and noise reductions are necessary Personalized service is more important
6
1. Motivation Large amount of images are available on Internet
Flickr.com is becoming very popular Hard for Query Formulation b. Information Overload
7
1. Motivation Applications Second-Generation Internet Search Engines
Family Photo Organization System Digital Library
8
2. Key Issues to be Addressed
How can we provide a good global overview of large-scale image collections such as Flickr images? How can we capture the users’ query intentions easily and effectively? How can we recommend the most relevant images according to the users’ intentions?
9
Why Flickr is attractive?
Low barrier to enter, low cognitive cost Immediate feedback and comments Individual & community aspects
10
What problems Flickr have?
Low Quality of Tagging: people may make mistakes Homonymy problem: one text term may have different means under different contexts Synonymy problem: many text terms may have same mean Tag redundancy
11
2. Key Issues to be Addressed
Query Recommendation: picture owners may use different text terms for image annotation, and users may not know these text terms Information overload: each text term may relate to many images in Flickr such as Roma
12
What Flickr has done for these?
Tagcloud for query recommendation
13
Problems for Tagcloud Only the frequency of text terms is used to characterized their importance, but the importance of the text terms may also depend on other issues Contexts between the text terms are missed, thus it cannot support effective navigation and exploration Such tagcloud may not be able to represent the users’ image needs effectively
14
Some improvements on tag vis.
15
Some improvements on tag vis.
16
Some improvements on tag vis.
17
Social Network Visualization
How to identify such contexts for social tags?
18
Our Proposed Solutions
Topic Network for providing a good global overview of large-scale image collections; Interactive Visualization for capturing the users’ query intentions easily and effectively; Interestingness Matching for recommending the most relevant images according to the users’ intentions.
19
3. Topic Network Construction
What can topic network represent? Global overview of large-scale image collections; Context among the topics; local contexts + global contexts Good navigation or browsing structure for accessing large-scale image collections.
20
3. Topic Network Construction
Major steps for concept network construction Natural language processing for entity extraction; Word frequency estimation; Keyword extraction; Co-occurrence for context determination.
21
3. Topic Network Construction
Social Tags & Discussions in Texts Social Tag Cleansing Topics & Topic Correlations Topic Network
22
3. Topic Network Construction
Social Tags & Discussions in Texts Social Tag Cleansing Keyword Extraction Correlation Analysis Topics & Topic Correlations
23
3. Topic Network Construction
Social Tags & Discussions in Texts Social Tag Cleansing Spam tags Noisy tags ……….. Topics & Topic Correlations
24
3. Topic Network Construction
Social Tags & Discussions in Texts Keyword Extraction Numbers of images Numbers of discussions Numbers of clicks ………. Topics & Topic Correlations
25
3. Topic Network Construction
Co-Occurrence Probability Semantic Similarity Distribution Similarity
26
3. Topic Network Construction
Context between concepts
27
3. Topic Network Construction
28
3. Topic Network Construction
29
4. Topic Network Visualization
Interestingness of topic self-significance: numbers of images, discussions & clicks Link-significance: numbers of linkages
30
4. Topic Network Visualization
Using interestingness to highlight image topics
31
4. Topic Network Visualization
Constraint-based topic clustering
32
4. Topic Network Visualization
Constraint-based topic clustering
33
4. Topic Network Visualization
Query-Driven Topic Network Visualization
34
4. Topic Network Visualization
Query-Driven Topic Network Visualization
35
4. Topic Network Visualization
Multidimensional Scaling
36
4. Topic Network Visualization
37
4. Topic Network Visualization
Change of Focus
38
4. Topic Network Visualization
Change of Focus---Poincare Mapping
39
4. Topic Network Visualization
Change of Focus---Poincare Mapping
40
5. Interactive Query Specification
Concept Network Navigation
41
5. Interactive Query Specification
Personalized concept network generation
42
5. Interactive Query Specification
43
6. Most Representative Image Recommendation
Image Content Representation
44
6. Most Representative Image Recommendation
Image Similarity Characterization Color Histogram Kernel
45
6. Most Representative Image Recommendation
Image Similarity Characterization Wavelet Filter Bank Kernel
46
6. Most Representative Image Recommendation
Image Similarity Characterization Interest Point Matching Kernel
47
6. Most Representative Image Recommendation
Image Similarity Characterization ---- mixture of kernels Multiple kernel learning for determining weighting parameters!
48
6. Most Representative Image Recommendation
Kernel-Based Image Clustering
49
6. Most Representative Image Recommendation
Support Vector Machine positive images negative images positive support vectors margin negative support vectors
50
6. Most Representative Image Recommendation
Representativeness of image
51
6. Most Representative Image Recommendation
Most Representative Image Selection Most representative images come from all these clusters and distribute according their coverage percentages; Images close to decision boundaries; Most frequent accessed images
52
6. Most Representative Image Visualization
Kernel PCA Similarity-Preserving Image Projection
53
6. Most Representative Image Visualization
Hyperbolic Image Visualization
54
6. Most Representative Image Visualization
55
6. Most Representative Image Visualization
56
6. Most Representative Image Visualization
57
7. User-Adaptive Image Recommendation
User Intention Capturing
58
7. User-Adaptive Image Recommendation
Personalized Interestingness of Image
59
7. User-Adaptive Image Recommendation
60
7. User-Adaptive Image Recommendation
61
Collaborative Filtering
Social Rating Information User’s Personal History Potential Changes over time Social Network
62
Multi-Modal Deep Learning
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.