Beyond Classification: Latent User Interests Profiling ...ylongqi/presentation/YangHE15Slides.pdfBeyond Classification: Latent User Interests Profiling from Visual Contents Analysis

Beyond Classification: Latent User Interests Profiling from Visual Contents Analysis

Longqi Yang, Cheng-Kang (Andy) Hsieh, Deborah Estrin

Social Network Online Purchases Communications

Our interests are manifested online …

Posted/Shared Contents People Connected/Followed Items Purchased

Preferences learning using small data

Online Posts

Private Communication

Shared Images

Personal Image Gallery

Preference Profile

News Search Engine Dietary Entertainment

Text/label-centric approach is widely studied

River

restaurant

tourism

landscape

Topic ModelingStructure Prediction

Classification/Labeling/Image-to-text

Travel Animal Art

But preferences are sometimes not just about text...

Intra-categorical variance: Hard to capture with text/label!

User A

User B

TravelImages

Research question

Images’ predictive power for users’ preferences beyond labels

Task 1: Pairwise ComparisonTask 2: Prediction

Pairwise Comparison

User A

User B

Discriminative Power of images

IMG1 IMG2 IMGn

IMG1 IMG2 IMGm

…...

…...

Same Label

Prediction: Consistency of Preferences

User 1

User N

Timeline

Predict/Retrieve

IMG1 IMG2 IMGn

IMG1 IMG2 IMGm

…... …...

…... …...

Same Label

Dataset

Travel boards

Background corpus Analysis

1,800 3,990

5,790 Travel boards

❶ ≥ 100 pins

❷ ∃ pins after June 2014

User Modeling and Image Representation

Pretrained Siamese Network

Pretrained Places CNN

205 d

im 410 dim205

dim

Pretrained cluster centers (200) 200 dim

…

User Profile

…

pins

B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva. “Learning Deep Features for Scene Recognition using Places Database.” Advances in Neural Information Processing Systems 27 (NIPS), 2014

*

*

User Modeling and Image Representation

Imag

e 2

Imag

e 1

x

y

A (CNN)

B (CNN)

Con

tras

tive

Los

s

f(x)

f(y)

− ≈ 0

− > 𝑚

,

,

𝓛 =𝟏𝟐𝒍𝑫

𝟐 +𝟏𝟐 𝟏 − 𝒍 𝐦𝐚𝐱(𝟎,𝒎− 𝑫)𝟐

𝒍 = 𝟏

𝒍 = 𝟎

Pairwise Comparison

User A

User B Effects of background distribution!

IMG1 IMG2 IMGn…...

Travel Images

IMG1 IMG2 IMGm…...

Pairwise Comparison

Document 1

Document 2

“and” 10%

“and” 11%

“fatuous” 0.001%

“fatuous” 1.001%

1% 1%

Background “and” 11% “fatuous” 0.001%

User A User B

Pairwise Comparison

Background corpus

𝛿9:;< 𝜎>(𝛿9:;<)Log-odds-ratio Uncertainty

𝑧9:;< =𝛿9:;<

𝜎>(𝛿9:;<)

Pairwise Comparison

Confidence)Level:)95%

Confidence)Level:)99%

𝐦𝐚𝐱 𝒛𝒌𝑨;𝑩 For all user pairs among 3,990 boards

Prediction

User 1

User N

Timeline

Sampled 100 pins

50 pins for test10~50 pins for train

IMG1 IMG2…... …...IMG51 IMG100

IMG1 IMG2…... …...IMG51 IMG100

Prediction

𝑴𝑹𝑹 =𝟏𝑵G

𝟏𝒓𝒂𝒏𝒌𝒊

𝑵

𝒊L𝟏

Conclusion

Online Posts

Private Communication

Shared Images

Personal Image Gallery

Preference Profile

Small data fueled preferences learning – what can we do next?

v Utilities of images beyond text/labels.

v Multi-modal data fusion

v End-to-end learning

http://www.cs.cornell.edu/~ylongqi

http://smalldata.io/

@ylongqi

[email protected]

For more information