Transcript
Page 1: The Machine Learning Guide to Fine Dining

1

The Machine Learning Guide to Fine Dining

Rani Nelken

Page 2: The Machine Learning Guide to Fine Dining

2

Motivation

Human decision making

Machine Learning

Page 3: The Machine Learning Guide to Fine Dining

3

How to pick the best restaurant?

Trust your server’s recommendations?

Stick with a favorite or try a new dish?

Page 4: The Machine Learning Guide to Fine Dining

4

Best restaurant for a group of friends?

Page 5: The Machine Learning Guide to Fine Dining

5

Ensemble methods and rank aggregation

Group restaurant

choice

ElectionsMeta-searchCombining classifiers

Page 6: The Machine Learning Guide to Fine Dining

6

Borda count: consensus over majority

Candidate John Paul George

1st Choice A A C2nd Choice B C A3rd Choice C B B

A: ASTAB: Grill 23 & BarC: Craigie on Main

Page 7: The Machine Learning Guide to Fine Dining

7

Borda count

Restaurant John Paul George Score

1st Choice A A C 32nd Choice B C A 23rd Choice C B B 1

Candidate ScoreA 3+3+2=8C 1+2+3=6B 2+1+1=4

Page 8: The Machine Learning Guide to Fine Dining

8

Beyond Borda Count

• Partial lists

• Uneven comparisons

• Enhanced Heuristics

• Efficiency

Page 9: The Machine Learning Guide to Fine Dining

9

More holistic view: Markov Chains

AB C

Edges represent preference

Self-edges

Nodes representrestaurants

Find stationary distribution using power method

Page 10: The Machine Learning Guide to Fine Dining

10

How to pick the best restaurant?

Stick with a favorite or try a new dish?

Trust your server’s recommendations?

Page 11: The Machine Learning Guide to Fine Dining

11

Bayesian classification

Trust server’s

recs?

Document classification

Page 12: The Machine Learning Guide to Fine Dining

12

Flip the problem using Bayes’ rule

• Instead of

estimate

• Reminder:

Page 13: The Machine Learning Guide to Fine Dining

13

Naïve Bayes

• How to estimate?

• Reduce to liking individual ingredients

Page 14: The Machine Learning Guide to Fine Dining

14

What about unknown ingredients?

Page 15: The Machine Learning Guide to Fine Dining

15

Solution: Laplace smoothing

http://www.youtube.com/watch?v=iGPldwfoddw

Page 16: The Machine Learning Guide to Fine Dining

16

How to pick the best restaurant?

Trust your server’s recommendations?

Stick with a favorite or try a new dish?

Page 17: The Machine Learning Guide to Fine Dining

17

Multi-arm Bandits

When to choose a

new dish?

Website optimization, Clinical trials

Page 18: The Machine Learning Guide to Fine Dining

18

ε-Greedy

ExploitFavorite

Usually

With low probability

ExploreNew

New dish 1

New dish 2

Page 19: The Machine Learning Guide to Fine Dining

19

How to choose between unknown dishes?Information used

Sophistication

ε-greedy

SoftmaxDishes’ previous success ratio

Dishes’ #triesUCB

GLM UCBModel of unseen dishes

Contextual bandits

Bayesian sampling

Page 20: The Machine Learning Guide to Fine Dining

20

Strategy for unknown dishes

• Prior estimate for based on ingredients

• Optimistic correction

Page 21: The Machine Learning Guide to Fine Dining

21

Life lessons from bandits

• Optimism in the face of uncertainty

• Minimize regret relative to other strategies

Page 22: The Machine Learning Guide to Fine Dining

22

Summary

Human decision making

Machine Learning

Page 24: The Machine Learning Guide to Fine Dining

24

Thank you

@RaniNelkenRani Nelken