Ratings in Distributed Systems: A Bayesian Approach Ratings in Distributed Systems: A Bayesian Approach

<-Back

```@InProceedings{mui-2001a,
author         = {Lik Mui and Mojdeh Mohtashemi and Cheewee Ang and Peter Szolovits and Ari Halberstadt},
title          = {Ratings in Distributed Systems: A Bayesian Approach},
year           = {2001},
review-dates   = {2004-08-08, 2004-08-09},
value          = {ba},
booktitle      = {Workshop on Information Technologies and Systems},
hardcopy       = {yes},
key            = {mui-2001a}
}
```

### Summary

Introduces "Bayesian probabilistic framework" for trust to include contextualized and personalized ratings. The context issue though raised isn't really addressed other than as a subscript and the assumption that all the math takes place within the framework of the context. It's essentially relegated to follow-up work.

There are agents and objects and boolean ratings and encounters.

There are attributes (countably infinite) which describe context in environment. Context is defined in terms of applicability of a set of attributes.

Reputation is R: AxAxC -> [0,1]

Update rule is newstate: RxD -> R.

Although authors make an effort to define trust and reputation, they then blur trust and reputation while using the model only for local reputation inference (based on an individual's encounters with another agent, without third party reporting.)

[ ] It's not clear to me at this time how we get from the priors of uniform distribution and then blend into the formulation to known strangers along a continuum, as in one case we are using the priors and another we are using all the encounters along a reputation chain.

Indirect trust (reputation) is modelled by recursively applying rules, but when there are parallel streams of info on an individual agent, the assertions are proportionally weighted along a scheme that uses Chernoff bounds to provide an upper limit of weighting and then pro-rates those not meeting the bound.

### Key Factors

How placed in context with other work:
• zacharia-1999a
• glass-2000a
• yu-2000a
• rouchier-2001a
• sabater-2001a
• esfandiari-2001a
• eBay
• Beta Distribution (dudewicz-1988a) [need to follow-up]
• Chernoff Bound (mui-2002b) [need to follow-up]
• mui-2001c (has experiments! cannot find!)

Problem Addressed: Informal nature of existing reputation calculation schemes.

Main Claim and Evidence: This is a better mathematical formulation for calculation of reputation; prior work relies mostly on intuitive appeal. Prior work too ad hoc.

Assumptions:

• Boolean ratings
• Beta distribution
• Ignorance of another modelled by uniform distribution
• No discount over time on influence of rating
• Assumption of approval rating being independent of other encounters (seems highly improbable).
• Cut-off on Chernoff bounds for weighting of evidence between multiple agents
• Chains of evidence use multiplicative combination.
• Full knowledge will be available, or at least quite a bit once encounters allow sharing. Perhaps other experiments round this out better.

Next steps:

• ! Extend from binary to generalized discrete or continuous ratings
• ! Procedure for inferring ratings for a context from others
• ! Using ontologies to relate contexts
• ! Using different ontological views to relate contexts
• Requirement to disclose personal rating information to others

Remaining open questions:

• (See next steps, above... though many of these don't appear to have been addressed)
• [ ] Does it work well? Can't see mui-2001c, mui-2002b has the model recapitulated, but no experiments, Thesis is skimpy on experiments using the probabilistic model as well.

### Quality

Originality is outstanding.
Contribution/Significance is excellent.
Quality of organization is outstanding.
Quality of writing is excellent.
<-Back