Rating? - ec.tuwien.ac.at

33
Questions Rating?

Transcript of Rating? - ec.tuwien.ac.at

Page 1: Rating? - ec.tuwien.ac.at

Questions

Rating?

Page 2: Rating? - ec.tuwien.ac.at

Experiment

Subjects reported…

« Too far from our location. »

« The place was closed that day. »

« I would have liked it if the tapa

were cold. »

Page 3: Rating? - ec.tuwien.ac.at

Set of

alternatives

Recommended

Item

Rating prediction

Collaborative Rating-based prediction

| r(

)

Page 4: Rating? - ec.tuwien.ac.at

Questions

Choice?

Page 5: Rating? - ec.tuwien.ac.at

Choice

Set

A

Chosen

alternative

Evaluation of

alternatives

CHOICE

Choice rule with preferences:

Rational choice

Page 6: Rating? - ec.tuwien.ac.at

Choice Rule with preferences

Axiom of utility theory

Choice rule with utilities:

Rational choice

Page 7: Rating? - ec.tuwien.ac.at

Choice

Set

A

Chosen

alternative

Evaluation of

alternatives

Utility1

Utility2

Utility3

UtilityN

CHOICE

Rational choice

Page 8: Rating? - ec.tuwien.ac.at

Definition of utility

Random utility models

- Vector of alternative’s attributes

- Vector of preferences on attributes

X1

X2

X3

X4

Decision-maker Item

Page 9: Rating? - ec.tuwien.ac.at

Random utility models

X1

X2

X3

X4

Decision-maker

perspective Item

Observed factors: Representative utility,

Unobserved factors: Random variable, ε

Researcher

perspective

Page 10: Rating? - ec.tuwien.ac.at

Choice Rule becomes probabilistic:

With probabilities:

Integrating over random variable ε, and assuming f as its density:

Random utility models

Page 11: Rating? - ec.tuwien.ac.at

If each ε is i.i.d, f(ε) is a Gumbel distribution :

Standard logit function:

Mixed logit model:

is considered a random variable with density g.

Standard and mixed logit models

Page 12: Rating? - ec.tuwien.ac.at

coefficients (preferences) of

Parameter estimation

1. Set of choices: Attributes per chosen alternative a` + Attributes of all

the alternatives a in choice set A.

2. Maximum likelihod procedures:

estimated by means of :

Page 13: Rating? - ec.tuwien.ac.at

Set of

alternatives

Recommended

Item

Rating prediction

Differences with rating-based predictions?

| r(

)

Page 14: Rating? - ec.tuwien.ac.at

Differences: concept of utility

Rational Choice

+

Random utility

Models

Collaborative

Rating-based

Models

Page 15: Rating? - ec.tuwien.ac.at

Differences: preference estimation

Rational Choice

+

Random utility

Models

Collaborative

Rating-based

Models

kNN:

Matrix factorization:

Estimated from r

Page 16: Rating? - ec.tuwien.ac.at

Differences: Researcher perspective

Rating prediction

Choice prediction

Deterministic Probabilistic

Page 17: Rating? - ec.tuwien.ac.at

Differences: the choice set

Page 18: Rating? - ec.tuwien.ac.at

Differences: the choice set

Page 19: Rating? - ec.tuwien.ac.at

Differences: the choice set

(Ariely, 2010)

Page 20: Rating? - ec.tuwien.ac.at

Experimental design

Santiago-Tapas Contest

Page 21: Rating? - ec.tuwien.ac.at

Experimental design

Page 22: Rating? - ec.tuwien.ac.at

Experimental design

Page 23: Rating? - ec.tuwien.ac.at

Participating restaurants 56

Different tapas offered 109

Tapas consumed 35000

Participants in the experiment 100

Participants complete experiment 18

Recommendations generated 500

Recommendations validated 81

Experimental design

Page 24: Rating? - ec.tuwien.ac.at

For each tapa, we gathered:

Choice sets

1. Area of tapa

2. Restaurant of tapa

Attributes

1. Type

2. Character

3. Restaurant

Rating

Dataset

Page 25: Rating? - ec.tuwien.ac.at

Choice-based models

1. Standard logit

2. Mixed logit

Baselines

1. Collaborative user-based (kNN)

2. Basic matrix factorization

Choice-based models and baselines

Page 26: Rating? - ec.tuwien.ac.at

Error Metric I: Given individual cn, true choice i and

recommended item j:

Error Metric II: Position of real choice in list of

recommendations:

Validation: 100 run, two-fold 75-25% training-test sets

validation, and leave-one-out validation

Evaluation

Page 27: Rating? - ec.tuwien.ac.at

Results: Models fitting

Page 28: Rating? - ec.tuwien.ac.at

Results: Prediction errors

Page 29: Rating? - ec.tuwien.ac.at

Results: Prediction errors

Page 30: Rating? - ec.tuwien.ac.at

Results: Prediction errors

Page 31: Rating? - ec.tuwien.ac.at

Discussion: Ratings useful in choice models?

Sure! Ratings can enhance the prediction power of choice-

based models!

Current research: Testing models with possitive results.

Page 32: Rating? - ec.tuwien.ac.at

Discussion: Real recommendations?

Project: Smart City Coruña

1. Goal: Event recommendations.

2. Technology: Recommendation Engine

Page 33: Rating? - ec.tuwien.ac.at

Paula Saavedra

Pablo Barreiro

Roi Durán

Rosa Crujeiras

(School of Mathematics)

María Loureiro

(School of Business)

Eduardo Sánchez Vila

[email protected]

CITIUS

Grupo de Sistemas Inteligentes

Universidad de Santiago de Compostela

Choice-based Recommender Systems