Recommender Systems

44
Recommender Systems Eric Nalisnick CSE 435

description

Recommender Systems. Eric Nalisnick CSE 435. …. How can businesses direct customers to groups of similar , interesting , relevant , and undiscovered items? . Recommender Systems!. Method #1: Memory-Based Collaborative Filtering. A. B. C. D. E. 0. 1. =. 0. C. 0. - PowerPoint PPT Presentation

Transcript of Recommender Systems

Page 1: Recommender  Systems

Recommender Systems

Eric NalisnickCSE 435

Page 2: Recommender  Systems
Page 3: Recommender  Systems

Page 4: Recommender  Systems

How can businesses direct customers to groups of similar, interesting, relevant, and undiscovered items?

Page 5: Recommender  Systems

Recommender Systems!

Page 6: Recommender  Systems

Method #1: Memory-Based Collaborative Filtering

Page 7: Recommender  Systems

A B C D E

Page 8: Recommender  Systems

=0100C

Page 9: Recommender  Systems

Pie Ice Cream Soup Egg Rolls

A 1 1 0 0

B 1 1 0 0C 0 1 0 0

D 0 0 1 1E 0 0 1 0

Customer—Item Matrix

Page 10: Recommender  Systems
Page 11: Recommender  Systems

= 00B

Page 12: Recommender  Systems

Pie Ice Cream Soup Egg Rolls

A 5 1 0 0

B 2 5 0 0C 0 4 0 0

D 0 0 3 3E 0 0 4 0

Sim.44

2.13-00

Customer—Item Matrix with User Reviews

Page 13: Recommender  Systems

Evaluation of Memory-Based Collaborative Filtering

Page 14: Recommender  Systems

1. Best for post-purchase recommendations.

Page 15: Recommender  Systems

2. Does not scale well.

Customers Items

Page 16: Recommender  Systems

3. Very popular and very unpopular items are problematic.

*In practice, can multiply values by inverse frequency

Page 17: Recommender  Systems

4. Cold Start Problem How do we recommend new items?

How do we make recommendations for new users?

Page 18: Recommender  Systems

5. Susceptible to Black and Gray Sheep

Page 19: Recommender  Systems

Method #2: Knowledge-Based Collaborative Filtering

Page 20: Recommender  Systems

Like traditional CBR systems…

Page 21: Recommender  Systems

Similarity function?

Page 22: Recommender  Systems
Page 23: Recommender  Systems

15

13

17

9

1

7

12

Page 24: Recommender  Systems

*Director, year, and color had unstable or negative weights.

Page 25: Recommender  Systems

Evaluation of Knowledge-Based Collaborative Filtering

Page 26: Recommender  Systems

1. Better at pre-purchase recommendations than Memory-Based.

Page 27: Recommender  Systems

2. Efficient runtime. Can be as simple as descending K-D Tree.

Page 28: Recommender  Systems

3. Cold Start problem and popularity of an item are not an issue.

Page 29: Recommender  Systems

4. Not good at modeling the general preferences of a user.

Page 30: Recommender  Systems

Method #3: Hybrid Item-to-Item Collaborative Filtering

Page 31: Recommender  Systems

A B C D E

Page 32: Recommender  Systems
Page 33: Recommender  Systems

Item-to-Item Collaborative Filtering AlgorithmFor each item i1:

For each customer c who has bought i1:

For each item i2 bought by c:Sim(i1, i2)

Page 34: Recommender  Systems

Pie Ice Cream Soup Egg Rolls

A 1 1 0 0

B 1 1 0 0C 0 1 0 0

D 0 0 1 1E 0 0 1 0

Customer—Item Matrix

Page 35: Recommender  Systems

Industry Example: The Netflix Prize

Page 36: Recommender  Systems

$1,000,000 prize

Page 37: Recommender  Systems

Winning Team: “Bellkor’s Pragmatic Chaos”

RMSE Reduction: 10.9%

Page 38: Recommender  Systems

Lessons Learned…

1. Baseline Predictors

Page 39: Recommender  Systems

Lessons Learned…

2. Binary view of Data: Rated or not rated.

Page 40: Recommender  Systems

Lessons Learned…

3. Restricted Boltzmann Machines.

Page 41: Recommender  Systems
Page 42: Recommender  Systems

Lessons Learned…

4. No one recommendation technique is best. Need to combine several.

Page 43: Recommender  Systems

Summary

1. Memory-Based CF is best for post-purchase

2. Knowledge-Based CF is best for pre-purchase.

3. Hybrid methods generally work best4. The data is as important as the

algorithm

Page 44: Recommender  Systems

Questions?