Relevant Reputation Protocol

rep:

detrep (deterministic version):

Abstract

Relevant Reputation Protocol is a personalized pagerank algorithm that supports negative links. It is used to compute user and content rankings in a reddit-like bulletin board — Relevant.

Disclaimer: The code in this repo has not be audited or tested in production, please use at your own risk.

Use Cases

Because the algorithm supports negative links, it can be used to represent upvotes and downvotes. This enables usescases such as voting, governance and ranking of data.

Deterministic Version

The deterministiv version of the algorithm uses Uint and safe-math libs from Cosmos Sdk to avoid floating point computations. It can be used in a blockchain environment where concensus is required.

Graph Visualizations:

https://relevant-community.github.io/reputation/

Usage

For complete usage examples, check out the test files: https://github.com/relevant-community/reputation/blob/main/rep/pagerank_test.go and https://github.com/relevant-community/reputation/blob/main/detrep/pagerank_test.go

As well as the R3l Cosmos App (WIP): https://github.com/relevant-community/r3l/blob/main/cmd/r3ld/cmd/worker/computeRep.go

Step 1. Import the Package and Initialize the Graph

import	rep "github.com/relevant-community/reputation/non-deterministic"

Create an instance of the graph:

graph := rep.NewGraph(0.85, 1e-8, 0)

NewGraph params:

α is the probability of not doing a jump during a random walk, usually .85, but can be 1 when including a personalization vector.
ε is the error margin required to approximate convergence, usually something small
The third parameter is a cached rank of the negative consumer node. This is an extra node we add to the graph to enable negative links, and this is its rank from previous computations.

Step 2. Add some Nodes

a := rep.NewNode("a", 0, 0)
b := rep.NewNode("b", 0, 0)
c := rep.NewNode("c", 0, 0)
d := rep.NewNode("d", 0, 0)
e := rep.NewNode("e", 0, 0)

Optionally indicate which nodes belong in the personalization vector:

graph.AddPersonalizationNode(a)

NewNode params:

NodeId - String
Positive Rank (cached from previous computation)
Negative Rank (cached from previous computation)

*Note: including cached values is necessary for computing the impact of negative links correctly and improves performance.

Step 3. Add Links between the Nodes

graph.Link(a, b, 2.0)
graph.Link(a, c, 1.0)
graph.Link(c, d, 1.0)
graph.Link(b, d, -1.0)
graph.Link(a, e, -1.0)

Link function creates a weighted directional link from nodeA to nodeB with a weight. Negative links have negative weight.

params:

ID of nodeA
ID of nodeB
Link weight (links are relative, so scale doesn't matter)

Step4. Run the Pagerank Algorithm

type Result struct {
	pRank float64
	nRank float64
}

result := map[string]rep.Result{}

graph.Rank(func(id string, pRank float64, nRank float64) {
  result[id] = rep.Result{
    pRank: pRank,
    nRank: nRank,
  }
})

The Rank method takes a callback parameter that will be called for each node.

Note: If you have negative links, you will want to take the results of the first pagerank computation, and run the algorithm again. This will ensure that the outgoing links from nodes that have a negative component caryy less weight.

Core Concepts and Features

Personalization

The personalization property of the graph designates trusted/authority nodes. These nodes will have a high reputation by default. Repuation flows from these nodes to the rest of the graph.

Personalized pagerank allows for construction of a sybil-resistant network. This means users that are not part of the personalization vector cannot manipulate the scores by creating sybil nodes that link to (or upvote) one another.

Notes:

If the personalization vector is left empty, the algorithm will not be sybil resistant and malicious nodes will be able to manipulate the rankings.
Nodes in the personalization vector have the power to manipulate rankings, so special care should be taken to when selecting them.

Negative Links

Negative links are possible because we represent each entity in our graph via two nodes - one positive and one negative. Negative links boost the ranking of the negative node, positive links boost the ranking of the positive node. These two scores can then be combined into a single reputation score.

However we cannot easily use the merged score inside the pagerank algorithm. When a node has both a positive and a negative score, only the positive node is used for in the computation, but we modulate the weight of its outgoing links based on the rank of the negative node.

This necesitates running the pagerank computation twice. The first round ensures we have computed negative and positive rankings of the nodes. The second round enables us to take the negative score into account and ignore nodes that have a high negative/positive rank ratio all together.

Implementation details: We modulate the weight of outgoing links by creating one global negConsumer node. Nodes that have both a negative and a positive rank, will have a portion of their outgoing weight consumed by a link to negConsumer, thereby decreasing the weight of other outgoing links.

TODOS:

Optimization & benchmarking - we should probably use int-indexed maps and fixed size arrays where possible.
Edge case (only impacts display) - if a node has no inputs we should re-set its score to 0 to avoid a stale score being displayed after all links to the node were removed or cancelled-out.

Credits

This pagerank implementation was used as a starting point. https://github.com/alixaxel/pagerank

relevant-community / reputation