user modeling and recommender systems: recommendation algorithms

34
User Modeling and Recommender Systems: recommendation algorithms Adolfo Ruiz Calleja 04/10/2014

Upload: ashlynn-hudson

Post on 18-Jan-2018

219 views

Category:

Documents


0 download

DESCRIPTION

Index Introduction Non-personalized recommender algorithms Content-based recommender algorithms Collaborative recommendation algorithms

TRANSCRIPT

Page 1: User Modeling and Recommender Systems: recommendation algorithms

User Modeling and Recommender Systems: recommendation

algorithmsAdolfo Ruiz Calleja

04/10/2014

Page 2: User Modeling and Recommender Systems: recommendation algorithms

Index

2

• Introduction• Non-personalized recommender algorithms• Content-based recommender algorithms• Collaborative recommendation algorithms

Page 3: User Modeling and Recommender Systems: recommendation algorithms

Index

3

• Introduction• Non-personalized recommender algorithms• Content-based recommender algorithms• Collaborative recommendation algorithms

Page 4: User Modeling and Recommender Systems: recommendation algorithms

Introduction: Added value of the Recommender Systems

4

• Provision of personalized recommendations

• Allows to persuade each customer with personalized information

• Serendipitous discovery

• Enables to deal with the long tail

Page 5: User Modeling and Recommender Systems: recommendation algorithms

Introduction: Recommender system schema

5

USER ITEM

Algorithm

rating

Set of user attributes

Set of user attributesSet of user

attributesSet of user attributesSet of user

attributesSet of user attributes

Set of user attributesSet of user

attributesSet of user attributesSet of item

attributes

Page 6: User Modeling and Recommender Systems: recommendation algorithms

Introduction: Predictions and recommendations

• Outputs of recommender systems

• Prediction ≈ how much a user would like an item– Numeric scored related to the predicted opinion of

the user about a specific item

• Recommendations ≈ suggestion of things you may like– It is typically a list of items– Internally has to make some predictions

Page 7: User Modeling and Recommender Systems: recommendation algorithms

Introduction: Proceed with caution

7

Page 8: User Modeling and Recommender Systems: recommendation algorithms

Index

8

• Introduction• Non-personalized recommender algorithms– Simple mean– Probabilistic method

• Content-based recommender algorithms• Collaborative recommendation algorithms

Page 9: User Modeling and Recommender Systems: recommendation algorithms

9

USER ITEM

Algorithm

rating

Set of user attributes

Set of user attributesSet of user

attributesSet of user attributesSet of user

attributesSet of user attributes

Set of user attributesSet of user

attributesSet of user attributesSet of item

attributes

Not personalized recommender algorithms

Page 10: User Modeling and Recommender Systems: recommendation algorithms

Not personalized recommender algorithms

10

• Based on External Community Data• Can know ephemeral information from the user

• Example: Tripadvisor or Booking

Page 11: User Modeling and Recommender Systems: recommendation algorithms

11

Simple mean

Page 12: User Modeling and Recommender Systems: recommendation algorithms

12

Probability method

Page 13: User Modeling and Recommender Systems: recommendation algorithms

Not personalized recommender algorithms

13

• Very simple algorithms• They forget about the long tail

• When there are lot of raters predictions tend to median score– Self-selection bias– Diversity of raters

• Pretty bad accuracy

Page 14: User Modeling and Recommender Systems: recommendation algorithms

Index

14

• Introduction• Non-personalized recommender algorithms• Content-based recommender algorithms– Explicit decision model– The vector space model

• Collaborative recommendation algorithms

Page 15: User Modeling and Recommender Systems: recommendation algorithms

Content-based recommendation

15

USER ITEM

Algorithm

rating

Set of user attributes

Set of user attributesSet of user

attributesSet of user attributesSet of user

attributesSet of user attributes

Set of user attributesSet of user

attributesSet of user attributesSet of item

attributes

Page 16: User Modeling and Recommender Systems: recommendation algorithms

Content-based recommendation

16

• User model is built analyzing user preferences and item attributes

• Hard to found massively used examples– Personalized news feeds

Page 17: User Modeling and Recommender Systems: recommendation algorithms

Explicit decision model

17

Page 18: User Modeling and Recommender Systems: recommendation algorithms

Explicit decision model

18

• Very well known method in many domains

• The decision tree can be automatically built– No need to formalize domain knowledge

• Can be used with small numbers of features– But recommender systems typically need very many

• They are almost never used

Page 19: User Modeling and Recommender Systems: recommendation algorithms

The vector space model

19

Page 20: User Modeling and Recommender Systems: recommendation algorithms

The vector space model

20

• Which factors to consider in the item description?

• Possibility to use keyword vector– It can be automatically extracted from text

• But not only for textual items!!– We can aggregate keywords

• But how?

• How to normalize the vector space?– Hard if it is not automatically done– Term Frequency-Inverse Document Frequency

• Do we trust on it?

Page 21: User Modeling and Recommender Systems: recommendation algorithms

The vector space model

21

• How to build the user profile?

• If I like it, it is important for me– Sometimes something I do not like may be relevant or

viceversa

• Problem of how to update user profiles– Are new items more important than previous ones?• Short term vs. Long term

Page 22: User Modeling and Recommender Systems: recommendation algorithms

The vector space model

22

• We do not need lot of users• Easy to compute and simple to implement• Flexible– Easy to integrate with other approaches– Quickly adapt to changes

• :S Hard to find out the factors and their weights• Cannot deal with subjective aspects of the items• Competitor items are frequently retrieved• Too simplified model– Results are not accurate as with other approaches

Page 23: User Modeling and Recommender Systems: recommendation algorithms

Index

23

• Introduction• Non-personalized recommender algorithms• Content-based recommender algorithms• Collaborative recommendation algorithms– User-based nearest neighbor recommendation– Item-based nearest neighbor recommendation

Page 24: User Modeling and Recommender Systems: recommendation algorithms

Collaborative recommendation algorithms

24

USER ITEM

Algorithm

rating

Set of user attributes

Set of user attributesSet of user

attributesSet of user attributesSet of user

attributesSet of user attributes

Set of user attributesSet of user

attributesSet of user attributesSet of item

attributes

Page 25: User Modeling and Recommender Systems: recommendation algorithms

Collaborative recommendation algorithms

25

• Item model is a set of ratings• User model is a set of ratings

• Predominant paradigm

Page 26: User Modeling and Recommender Systems: recommendation algorithms

User-based nearest neighbor recommendation

26

Page 27: User Modeling and Recommender Systems: recommendation algorithms

27

• Pearson correlation coefficient• There are other algorithms– But commonly provide less accurate results– Cosine correlation is becoming on fashion

• Pearson correlation has some deficiencies– What if two users have few items in common?– What if the ratings are unary data?– What if something is loved or hated by the whole

community?

User-based nearest neighbor recommendation

Page 28: User Modeling and Recommender Systems: recommendation algorithms

28

• Processing time = O(N^2*M)– But not in real life

• Neighborhood selection– 20 to 50 neighbors (sometimes up to 100)– Define number of neighbors or a threshold– Better processing time O(N*M)– Less noise– Reduce coverage

User-based nearest neighbor recommendation

Page 29: User Modeling and Recommender Systems: recommendation algorithms

29

• Precomputed neighborhood– Better response time– Need to be frequently update (it is not a good idea to

define clusters)

User-based nearest neighbor recommendation

Page 30: User Modeling and Recommender Systems: recommendation algorithms

30

• Very popular• Based on subjective information• Very many variants and possible configurations

• What do we do with new items?• What do we do with new users?• Need of (similar) users• Data sparcity is a problem

User-based nearest neighbor recommendation

Page 31: User Modeling and Recommender Systems: recommendation algorithms

Item-based nearest neighbor recommendation

31

Page 32: User Modeling and Recommender Systems: recommendation algorithms

32

• Pearson correlation coefficient or cosine similarity– But now the neighborhood is formed by items!!

• A model should be built– Processing time = O (I^2)– It is always precomputed– Do not need to save all the model• Memory used vs. accuracy and coverage

– Items are much more stable that users• But they still need to be updated

Item-based nearest neighbor recommendation

Page 33: User Modeling and Recommender Systems: recommendation algorithms

33

• Efficient algorithm• Scales very well• Data sparcity is not a big problem• Creates nice recommendation lists

• We still need to deal with the cold-start• Memory use

Item-based nearest neighbor recommendation

Page 34: User Modeling and Recommender Systems: recommendation algorithms

User Modeling and Recommender Systems: recommendation

algorithmsAdolfo Ruiz Calleja

04/10/2014