A Generalized Stereotypical Trust Model - Nanyang Technological ...

Motivation RelatedWork Approach Experiment Conclusion

A Generalized Stereotypical Trust Model Hui Fang, Jie Zhang, Murat S¸ensoy, and Nadia Magnenat-Thalmann {[email protected]} IMI & SCE, Nanyang Technological University, Singapore

August 14, 2012

.

H. Fang, J. Zhang, M. S ¸ ensoy, N. Thalmann

.

.

.

A Generalized Stereotypical Trust Model

.

.

1 / 21


Outline

1 Motivation & Objectives 2 Related Work 3 Approach: Generalized Stereotypical Trust Model 4 Experimentation 5 Conclusions and Future Work

.


.

.

.


.

.

2 / 21


Motivation & Objectives

Model trustworthiness of sellers: Experienced sellers Sufficient past experience Reputation systems

Inexperienced sellers No or little past experience Stereotypical trust models

.


.

.

.


.

.

3 / 21



Stereotypical trust model for a buyer: Basic assumptions Seller behavior follows certain patterns Evaluation on a new seller influenced by past experience

Machine Learning on his own experience Cold start problem Insufficient past experience with sellers Can not learn accurate trust stereotypes Collect other buyers’ experience Users’ subjectivity difference

.


.

.

.


.

.

4 / 21



A generalized stereotypical trust model for cold start problem: Build a semantic ontology Represent sellers’ attributes in e-marketplaces

Fuzzy semantic decision tree (FSDT) with limited experience Fuzzy process: generalize over non-nominal attributes Semantic process: generalize over nominal attributes

.


.

.

.


.

.

5 / 21


Related Work Stereotypical trust models: Burnett et al. (2010) [1] Use M5 tree to learn stereotypes Integrate with probabilistic trust model

Liu et al. (2009) [2] Apply traditional machine learning tools Directly predict successfulness of transactions

Collect other buyers’ experience for cold start problem Limitations Users’ subjectivity difference problem Manually identify attributes

Our approach Only use a buyer’s own experience Ontological reasoning .


.

.

.


.

.

6 / 21


Related Work Stereotype-based user modeling, e.g.: Rich (1979) [3] Manually construct users’ stereotypes User’s socio-demographic characteristics

Ardissono et al. (2003) [4] Personalized electronic guides for digital TV Stereotypical user model with machine learning

Limitations Focus on socio-demographic characteristics Cold start problem

Our approach Also consider attributes related to sellers’ selling behavior Address cold start problem .


.

.

.


.

.

7 / 21


Semantic Ontology .

Represent the hierarchical relationships among attribute values Automatically identify seller attributes and values Ontological reasoning technique helps our semantic process Thing isA

isA

Seller

hasProperty

Property isA

isA

sellingProperty

socio-demographicProperty isA

gender

isA

isA

systemAge

...

price

location

ocean

isA

category

deliveryType

...

continent Africa China

Asia Singapore

Japan

Europe Spain

UK

Figure: An Example Semantic Ontology of Seller Attributes in E-marketplaces

.

.


.

.

.


.

.

8 / 21


Fuzzy Semantic Decision Tree Learning

Goal of stereotypical learning

− → F : A → Ts − → A : attribute vector describing seller information Ts : the trust degree (ranged in [0, 1]) for the seller

Our approach: FSDT to learn function F Basic decision tree learning + two additional processes Fuzzy process for non-nominal attributes Semantic process for nominal attributes

.


.

.

.


.

.

9 / 21


Fuzzy Process Generate fuzzy decision tree (Figure 2(a)): Extend the fuzzy method [5] with non-nominal attributes Piecewise Linear Membership function (Figure 2(b)) Instances in [α − β/2, α + β/2] to both the successors

Traditional decision tree with nominal attributes T1: 100 0.66 Singapore T2: 70

α

location β

0.88

α-β/2 L2: 66

110

L3: 10

UK

L1: 30 0.00

α+β/2

0.1

(a) A Fuzzy Decision Tree Example

Piecewise linear 1 0.5 0

(b) Piecewise Linear Membership .


attribute

α β

.

.

.


.

.

10 / 21


Fuzzy Process Split Criterion: minimize the mean square error (MSE) ∑ ˆ L )2 MSE(S, a) = mean[ µS (i) × (Ti − T i∈SL

+

∑

(3.1)

ˆ R )2 ] µS (i) × (Ti − T

i∈SR

MSE(S, a) = mean[

∑

ˆ i )2 ] µS (i) × (Ti − T

(3.2)

i∈S

ˆL + [1 − µ(i, a, α, β)] × T ˆR ˆi = µ(i, a, α, β) × T T

(3.3)

where a is a certain attribute; Learn (α, β) for each non-nominal attribute a .


.

.

.


.

.

11 / 21


Fuzzy Process

Contribution of fuzzy process: Improve performance of stereotypical learning Intuition Both groups with corresponding membership degree Fuzzy area With limited experience Difficult to precisely classify inexperience seller

.


.

.

.


.

.

12 / 21


Semantic Process Based on fuzzy decision tree: Exploit hierarchial relationships among attribute values Generalize over the values of nominal attributes Replace specific values with more general one Without significant decrease of classification performance E.g, a location ontology Thing DeliveryType

...

Location

Continent Asia

Africa China

Ocean

Singapore

Europe

Japan

UK

Spain

Figure: A Location Ontology .


.

.

.


.

.

13 / 21


Semantic Process

T1: 100

T1: 100 0.66

0.66 Singapore location T2: 70

α

β

0.88

α-β/2 L2: 66

110

L3: 10

L1: 30 0.00

α+β/2

T2: 70

α

(a) Fuzzy Decision Tree

Location

110

0.98

0.1

L2: 66

L3: 10

L1: 30 0.00

(b) Fuzzy Semantic Decision Tree

Contribution of semantic process Can predict the trustworthiness of inexperienced sellers with unmet values for nominal attributes .


.

.

.


.

.

14 / 21


Simulation Simulated e-marketplace: 100 sellers and 20 buyers Sellers provide products of different quality Five attributes: location, system age, price, delivery type and number of items sold Buyers have different subjectivity

Benchmark Comparisons Baseline: random with memory SR: Burnett et al.’s approach [1] FUZZY: only fuzzy process SEMANTIC: only semantic process

Parameters Degree of buyer subjectivity difference Value ranges of non-nominal attributes The number of possible values of nominal attributes .


.

.

.


.

.

15 / 21


Experimental Results Our model performs better than SR and Baseline (Figure 3(c)) Both the fuzzy and semantic process contribute to performance improvement (Figure 3(d)) FUZZY and SEMANTIC outperform the SR approach (Figures 3(c) and 3(d)) 0.7

0.7 Our Model SR Baseline

Our Model SEMANTIC FUZZY

0.6

0.5

0.5

0.4

0.4

MAE

MAE

0.6

0.3

0.3

0.2

0.2

0.1

0.1

0

0 0

10

20

30

40

50 60 Epoch

70

80

90

(c) FSDT vs. SR vs. Baseline

100

0

10

20

30

50 60 Epoch

70

80

90

100

(d) FSDT vs. FUZZY vs. Semantic .


40

.

.

.


.

.

16 / 21


Experimental Results Worse than Reputation-based method and SR when |d| is small, better when |d| gets larger Our Model SR Reputation

0.14 0.12

MAE

0.1 0.08 0.06 0.04 0.02 0 0

0.2

0.4

0.6

0.8

1

|d|

Figure: Varying degree of buyer subjectivity difference

.


.

.

.


.

.

17 / 21


Experimental Results Figure 4(a) FUZZY performs better than SR Verify the generalization capability of the fuzzy process

Figure 4(b) SEMANTIC performs better than SR Verify the generalization capability of the semantic process The generalization capability increases as the number increases 0.3

0.3 FUZZY SR Performance Gap

0.25

0.25 0.2 MAE

MAE

0.2 0.15

0.15

0.1

0.1

0.05

0.05

0

SEMANTIC SR Performance Gap

0 0

5 10 15 20 25 Value Range of Non-Nominal Attributes

30

0

5 10 15 20 25 30 Number of Possible Values for Nominal Attributes

(a) Varying the value ranges of non-(b) Varying the number of possible nominal attributes values of nominal attributes .


.

.

.


.

.

18 / 21


Conclusions

Contribution of our approach: Propose a novel generalized stereotypical trust model Fuzzy process on non-nominal attributes Semantic process on nominal attributes

Address the cold start problem for stereotypes learning When buyers have limited experience

Verify the effectiveness of FSDT by experiments More accurately evaluate trustworthiness of inexperience sellers Robust with regard to the change of training data set

.


.

.

.


.

.

19 / 21


Future Work

Conduct more experiments More scenarios in the simulated e-marketplaces Real data (e.g., eBay)

Use other membership functions e.g., triangular and trapezoidal membership function

Consider multi-nominal trust degrees

.


.

.

.


.

.

20 / 21


Thanks! Any Questions?

.


.

.

.


.

.

21 / 21


C. Burnett, T. J. Norman, and K. Sycara. Bootstrapping trust evaluations through stereotypes. Proceedings of the 9th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS), 2010. X. Liu, A. Datta, K. Rzadca, and E.P. Lim. Stereotrust: a group based personalized trust model. in Proceedings of the 18th ACM Conference on Information and Knowledge Management (CIKM), 2009. E. Rich. User modeling via stereotypes. Cognitive Science, vol. 3, no. 4, pp. 329 C354, 1979. L. Ardissono, C. Gena, P. Torasso, F. Bellifemine, A. Chiarotto, D. A., and B. Negro. Personalized recommendation of tv programs. in Proceedings of the 8th Advances in Artificial Intelligence, 2003. .


.

.

.


.

.

21 / 21


C. Olaru and L. Wehenkel. A complete fuzzy decision tree technique. Fuzzy Sets and Systems, vol. 138, pp. 221 C254, 2003.

.


.

.

.


.

.

21 / 21