Aggregation techniques

Main article: Technical overview of the ratings system

Aggregation refers to the way we combine the opinions of others to obtain a final value of the opinion. A poll which takes the sum of each person's candidate preference in an election and then calculates the percentage for each candidate is an aggregation technique.

A number of aggregation techniques are possible:

1. Bayes' equation with a simple example of its use.
2. Simple averaging, a privacy enhancing variant thereof, and an averaging technique using with continuous input distributions, complex trust, and intermediate results.
3. Trust weighted averaging
4. An analysis of these methods and simple weighted averaging
5. Trust-weighted histograms
6. Trust/Probability/Population graphs algorithm
7. Binned and continuous distributions
8. Population distributions and graphical output with privacy

Privacy is an important issue in ratings aggregation and each algorithm will require customized treatment to deal with it.

More algorithms will be developed over time. Furthermore, the software will be built with an API to allow users to add their own algorithms.

Aggregation does not refer just to algorithms. Aggregators will likely be people chosen to perform aggregation tasks. Many raters may want to preserve anonymity by having some users perform the function of aggregating ratings and making the aggregated ratings public. We have discussed above methods of passing ratings information along from node to node in a way that preserves the privacy of individual ratings. Nevertheless, it is highly probable that we will still need aggregators to solve this problem in a comprehensive way.

A person of high integrity and technical competence would be needed for this job. The community would be relying on them to preserve the anonymity of everyone else and correctly apply any needed aggregation algorithms. Much of this will likely be automatable but any algorithmic aggregation will have to be according to the policy of a community and the aggregator would likely be the person selected to ensure that.

In general we will have two aggregator models:

1. One based on the subjective ratings system where each rater picks an aggregator and tells it, in effect (using privacy preserving methods), who the raters are. The aggregator is essentially a math engine. Methods have been proposed for minimizing nefarious behavior such as collusion in an attempt to expose one person’s rating. One advantage of this system is the ability to query a different aggregator and expect the same result.
2. A public model more suited to the community ratings system (CRS). The aggregator has known raters (selected by the community) so they are not specified by the person asking the question. This aggregator node is public so one advantage is that people can find it. People trust that it’s a safe place to send their opinion. We could have more than one aggregator but then, to check the answer, we’d have to make sure they used the same sources. In this system we don’t necessarily know who the raters are. Therefore, we don’t have to worry about one possible attack: a person submitting a question who tries to isolate a single response opinion by submitting a trust vector of zero for everyone else. The trust vector is presumably built-in to the public system by the community.