Problem:
What & Where to eat?
Mei Gao
http://imfeelinghungryy.com/
Topic0
japanese
Topic1
mexican
Topic2
brunch
Topic3
bar/atmosphere
Topic4
pizza
Topic 5
compliment
sushi tacos breakfast great pizza great
roll mexican coffee beer crust best
pita salsa eggs happyhour wings love
tuna burrito bacon bar thin good
salmon chips pancakes drinks pepperoni like
Topic6
indian
Topic7
asian
Topic8
fastfood
Topic9
sweets
Topic10
bbq
Topic 11
service(bad)
indian thai burger bagels cheese service
buffet pho fries cheese bbq didn’t
masala chinese potato best sauce never
naan soup onion rings smoothies chicken even
bianco curry dog Iove ribs back
LDA (Latent Dirichlet Allocation) 12 topics
Good restaurant: average star>3.5 Bad restaurant: average star<=3.5
Classification : Weight for each topic
Classifier Linear SVM Logistic
Regression
Random Forest
Accuracy in
Cross Validation 73.67% 81.19% 77.7%
Very Deep learning for image ranking
Training
Evaluation of Recommendation Error Using
Normalized Distance-based Performance Measure (NDPM)
R_2
R_5
R_1
R_3
R_4
R_1
R_2
R_3
R_4
R_5
Recom Actual Ranking
(ground truth)
Minimize NDPM score
Goal:
Mei Gao
Very Deep learning for image ranking
Food Images
Training
Application
Deep hierarchical abstraction Learning structure of images
Deep learning for image ranking
Assessment of LDA
BOW (Bag of Words) LDA
Feature
Dimension 10000 words in dictionary 15 topics
>99%
dimension
reduction
Computation
Efficiency 2.5 hrs 15 min
>90%
computation
time
(2000 samples)
(10 fold cross validation)
Topic0 Topic1 Topic2 Topic3 Topic4
Japanese Mexican brunch
Bar/ pizza
Atmosphere
sushi tacos breakfast great pizza
roll mexican coffee beer crust
pita salsa eggs happyhour wings
tuna burrito bacon bar thin
salmon chips pancakes drinks pepperoni
Topic5 Topic6 Topic7 Topic8 Topic9
Indian Asian fastfood sweets bbq
Indian Thai burger bagels cheese
buffet Pho fries cheese bbq
masala Chinese potato best sauce
naan soup Onion ring smoothies chicken
bianco curry dog Iove ribs
LDA (Latent DiriChlet Allocation) 15 topics
Topic 10: Compliment
Great, best, live, good, like
Topic 11: Service ( Bad)
Service, didn’t, never, even, back
Assessment of LDA
Dimension Reduction: 99% reduction in dimension
BOW (bag of words) features: 10,000 LDA features: 15 Topics
Computation efficiency: 2000 samples, 10 fold cross validation
BOW features: 2.5 hrs LDA features: 15 min
3%
4%
10%
4%
5%
29%
20%
6%
3%
16%
Percentage
others Japanese Mexican Brunch
bar Service compliment Asian
fastfood bbq
Evaluation of Recommendation Error Using
Normalized Distance-based Performance Measure (NDPM)
Rank by
recommendation
Rank by user's
actual ratings
Restaurant_1 Restaurant_1
Restaurant_2 Restaurant_3
Restaurant_3 Restaurant_7
Restaurant_4 Restaurant_2
Restaurant_5 Restaurant_4
Restaurant_6 Restaurant_9
Restaurant_7 Restaurant_8
Restaurant_8 Restaurant_5
Restaurant_9 Restaurant_10
Restaurant_10 Restaurant_6

Mei gao practicedemo_5

  • 1.
    Problem: What & Whereto eat? Mei Gao
  • 2.
  • 3.
    Topic0 japanese Topic1 mexican Topic2 brunch Topic3 bar/atmosphere Topic4 pizza Topic 5 compliment sushi tacosbreakfast great pizza great roll mexican coffee beer crust best pita salsa eggs happyhour wings love tuna burrito bacon bar thin good salmon chips pancakes drinks pepperoni like Topic6 indian Topic7 asian Topic8 fastfood Topic9 sweets Topic10 bbq Topic 11 service(bad) indian thai burger bagels cheese service buffet pho fries cheese bbq didn’t masala chinese potato best sauce never naan soup onion rings smoothies chicken even bianco curry dog Iove ribs back LDA (Latent Dirichlet Allocation) 12 topics
  • 4.
    Good restaurant: averagestar>3.5 Bad restaurant: average star<=3.5 Classification : Weight for each topic Classifier Linear SVM Logistic Regression Random Forest Accuracy in Cross Validation 73.67% 81.19% 77.7%
  • 5.
    Very Deep learningfor image ranking Training
  • 6.
    Evaluation of RecommendationError Using Normalized Distance-based Performance Measure (NDPM) R_2 R_5 R_1 R_3 R_4 R_1 R_2 R_3 R_4 R_5 Recom Actual Ranking (ground truth) Minimize NDPM score Goal:
  • 7.
  • 8.
    Very Deep learningfor image ranking Food Images Training Application
  • 9.
    Deep hierarchical abstractionLearning structure of images Deep learning for image ranking
  • 10.
    Assessment of LDA BOW(Bag of Words) LDA Feature Dimension 10000 words in dictionary 15 topics >99% dimension reduction Computation Efficiency 2.5 hrs 15 min >90% computation time (2000 samples) (10 fold cross validation)
  • 11.
    Topic0 Topic1 Topic2Topic3 Topic4 Japanese Mexican brunch Bar/ pizza Atmosphere sushi tacos breakfast great pizza roll mexican coffee beer crust pita salsa eggs happyhour wings tuna burrito bacon bar thin salmon chips pancakes drinks pepperoni Topic5 Topic6 Topic7 Topic8 Topic9 Indian Asian fastfood sweets bbq Indian Thai burger bagels cheese buffet Pho fries cheese bbq masala Chinese potato best sauce naan soup Onion ring smoothies chicken bianco curry dog Iove ribs LDA (Latent DiriChlet Allocation) 15 topics Topic 10: Compliment Great, best, live, good, like Topic 11: Service ( Bad) Service, didn’t, never, even, back
  • 12.
    Assessment of LDA DimensionReduction: 99% reduction in dimension BOW (bag of words) features: 10,000 LDA features: 15 Topics Computation efficiency: 2000 samples, 10 fold cross validation BOW features: 2.5 hrs LDA features: 15 min
  • 13.
    3% 4% 10% 4% 5% 29% 20% 6% 3% 16% Percentage others Japanese MexicanBrunch bar Service compliment Asian fastfood bbq
  • 14.
    Evaluation of RecommendationError Using Normalized Distance-based Performance Measure (NDPM) Rank by recommendation Rank by user's actual ratings Restaurant_1 Restaurant_1 Restaurant_2 Restaurant_3 Restaurant_3 Restaurant_7 Restaurant_4 Restaurant_2 Restaurant_5 Restaurant_4 Restaurant_6 Restaurant_9 Restaurant_7 Restaurant_8 Restaurant_8 Restaurant_5 Restaurant_9 Restaurant_10 Restaurant_10 Restaurant_6