In the Search of Better Deals Using Trust Joana Urbano, Ana Paula - - PowerPoint PPT Presentation

• –

–

• In the Search of Better Deals Using Trust

Joana Urbano, Ana Paula Rocha, Eugénio Oliveira

LIACC, DEI / Faculdade de Engenharia, Universidade do Porto {joana.urbano, arocha, eco}@fe.up.pt

August 17th 2010, IAT4EB @ECAI 2010

• –

–

• Motivation
• Our Computational Trust Proposal
• Evaluation & Conclusions

Outline

• –

–

• T (A, B, context)

Trust reflects the expectation on the activities of an entity when it reacts on a given context (Dasgupta, 2000)

• Application of Computational Trust:

Computational Trust

• Application of Computational Trust:

– Social networks – Recommender systems – Usage of network resources (e.g. telco, grid) – Electronic business

• –

–

• For the automation of several electronic business

processes:

– Negotiation – Selection of partners

Trust as an Enabler Technology

– Adaption of contracts, norms and sanctions – Any form of collaboration

• –

–

• Main areas of research:

– Semantics of trust and reputation – Anti-fraud mechanisms – Aggregation of trust evidences for the estimation of the trustworthiness value of agents

Work on Computational Trust

• Traditional approaches in aggregation:

– Beta and Dirichlet distributions – Statistical simple means – Statistical weighted means (by recency, by confidence, …)

• –

–

• They do not differentiate between different deceptive events:

– Delay in delivery – Received quality (e.g. affordability, safety, degree of uniqueness) – Quantity – Violation in intellectual property rights

Traditional Approaches Are Not Contextual

– Violation in intellectual property rights – Ethical problems, legislations, price fluctuation, etc

Business players tend to behave parochially

• –

–

• We intend to enhance traditional aggregation

methods:

1. Using dynamics of trust 2. Taking into account the situational context

Motivation

3. Allowing for heterogeneous evidences 4. Using techniques that work with few trust evidences

• –

–

• We intend to build an agent-based simulation

environment that allows us to evaluate:

1. Whether the break of breeding business relations allows to increase utility of clients or jeopardize it

Motivation – 2

2. How different trust models support the exploration of new partners in a safe way

• –

–

• Our Proposal
• We propose:

– SinAlpha, an aggregation engine that embeds dynamics of trust properties

SinAlpha Contextual

tune

– Contextual Fitness, a component for situation-aware trust

Contextual Fitness

CTR

• –

–

• Scenario

CFP (fabric; quantity; delivery time) PROPOSE

trust

manufacturers & exporters client

Contractual History trust evidences

SELECTION OUTCOME (true/false)

• –

–

• Motivation for Contextual Fitness

fabric quantity

• del. time
• utcome

cotton medium big true cotton high medium true cotton medium big true voile medium medium true voile high low false cotton high medium true voile medium big true

This supplier generally fulfils its commitments … but, does it have any handicap? Non-contextual

voile medium big true voile high big true cotton low big true voile medium medium true voile medium big true voile medium medium true cotton low big true cotton high medium false voile medium low true voile medium big True voile medium low true

Non-contextual aggregation methods do not capture these context specificities

• –

–

• Motivation for Our Approach

fabric quantity

• del. time
• utcome

cotton medium big true cotton high medium true cotton medium big true voile medium medium true voile high low false cotton high medium true voile medium big true

This supplier generally fulfils its commitments … but, does is have any handicap? Traditional

We propose to extract behaviour tendencies in a

voile medium big true voile high big true cotton low big true voile medium medium true voile medium big true voile medium medium true cotton low big true cotton high medium false voile medium low true voile medium big True voile medium low true

Traditional aggregation methods do not capture those context specificities

tendencies in a

• nline, dynamic,

and incremental way

• –

–

• Information Gain Technique
• 1. Builds a tree for every new

evidence using the information gain metric

fabric quantity

• del. time
• utcome

cotton medium big true cotton high medium true cotton medium big true voile medium medium true voile high low false cotton high medium true voile medium big true

• 2. Derives a tendency of failure

cfp cfp cfp cfp = cotton, 1080000, 7 = cotton, 1080000, 7 = cotton, 1080000, 7 = cotton, 1080000, 7 (*, high, low) (*, high, low) (*, high, low) (*, high, low)

voile medium big true voile high big true cotton low big true voile medium medium true voile medium big true voile medium medium true cotton low big true cotton high medium false voile medium low true voile medium big True voile medium low true

trustA (B, context) = trustA (B) * CFA (B, context)

• –

–

• Experiments – Populations

Supplier Types

• Prob. Success

HFab, HQt, HDt HFabQt, HFabDt, HQtDt 0.05 (handicap) 0.95 (all other)

Suppliers are assigned different handicaps at setup

HFabQt, HFabDt, HQtDt 0.95 (all other)

We evaluate how different trust models allow for better knowledge of these suppliers’ handicaps

• –

–

• Experiments – First Set
• How different trust models

explore new business

• pportunities
• M. Rehak, M. Gregor, M. Pechoucek. “Multidimensional

context representations for situational trust” (2006)

• Approaches in evaluation:

The context space is a Q-dimensional metric space with one dimension per each represented situation feature

– TradCTR: SinAlpha – InfGain: Contextual Fitness with information gain metric – ContSpace: a situation-aware technique based on pre- defined similarity measures

• –

–

• Results – First Set

close to 95% (best case) utility different number of suppliers

• –

–

• SA tends to select known partners that occasionally fail a

contract

• CS needs several evidences to populate reference contexts,

which also privileges parochialism

• CF is able to extract tendencies with a reduced number of

Interpretation of Results – First Set

• CF is able to extract tendencies with a reduced number of

evidences → bad “exploration” decisions are easily detected

• CF does not need pre-defined domain knowledge to operate

• –

–

• Suppliers have an internal value

Experiments – Second Set

0.50, 0.60, 0.70, 0.80, 0.90

• Clients agents use CF and can be of two types:

– Parochial: select partners based on their estimated trustworthiness – Explorative: … also consider the estimated internal value utility = trustworthiness * internal value

• –

–

• Results – Second Set

successful contracts different suppliers utility contracts suppliers

• –

–

• The flexibility of the online tendency extraction allows to

safely exploring a larger space of opportunities, e.g. better prices

• Exploring outside the set of known trustworthy partners can

bring extra utility to client agents

Interpretation of Results – Second Set

bring extra utility to client agents

• –

–

• We develop a trust model approach based on the

information gain metric that is:

– Situation-aware – Reactive to the dynamics of the suppliers’ performance

Conclusions

– Reactive to the dynamics of the suppliers’ performance – Reliable even in the presence of few evidences – More suppliers’ exploration supportive

• –

–

• Thank you!

joana.urbano@fe.up.pt