kg-disinfo

Knowledge Graph-based (dis)credibility estimation.

Given a KG and a few seed nodes which contain (lack of) credibility scores (represented as a numerical value), the system uses a metric propagation algorithm to estimate the (lack of) credibility of neighbouring nodes (for which no previous score is available).

The algorithm allows you to define weights based on relation types: this allows you to define which relations should propagate more or less of the disinformation metric.

Installation

Check the releases and download the standalone jar. Otherwise, you need to build it using leiningen as discussed below.

Set up

The project is built using Leiningen, which is like maven, but for Clojure.

Usage

With standalone jar

If you already have a standalone jar for this project you can call

$ java -jar kg-disinfo-0.2.0-standalone.jar [args]

e.g.

$ java -jar kg-disinfo-0.2.0-standalone.jar --help 

will print the command line arguments. Similarly

$ java -jar kg-disinfo-0.2.0-standalone.jar --graph resources/dante-g-small.json --injections resources/dante-g-small-injection.json

will:

• read a graph from the specified path
• read an injection from the specified path
• execute the appleseed algorithm on the graph and the seed injection, using 0.9 as the spreading-factor
• write the resulting scored graph to C:/tmp/dante-g_scored.pdf

Using leiningen

Either 1. build the standalone jar by calling

$ lein uberjar

and follow instructions described above, or

2. Use the run option, e.g.:

   $ lein run -- --help

This will do the same as running it with the standalone jar as described above.

Via repl (or programammatically)

Please refer to the code for examples on how to do this.

Specifying graphs and injections

graphs

The easiest way to use kg-disinfo is to provide your graph in a json format. A (automatically generated) example is provided. The file must contain:

• a single javascript object with attribute edges, which must be a list of edge objects.
• each edge object has fields: ** src-node mandatory a string identifier for the source node ** tgt-node mandatory a string identifier for the target node ** meta metadata for the edge with required fields: *** weight: mandatory this is the propagation weight between the source and the target node and should be a float between 0.0 and 1.0. Values close to 0.0 mean that (lack of) credibility scores from the source node will not be propagated to target node. Values close to 1.0 mean that (lack of) credibility scores from the source node will be propagated to the target node. *** name: optional the name of the relation

Note that the graphs are directed, therefore, you may need to add inverse edges if appropriate. This is needed because (mis)credibility may propagate in one direction but not in the inverse.

Here's a short sample of what graph edges should look like.

{
  "edges" : [ {
    "src-node" : "digitalId_23585",
    "tgt-node" : "person_23578",
    "meta" : {
      "weight" : 1.0,
      "name" : "isOwner-1"
    }
  }, {
    "src-node" : "digitalId_23585",
    "tgt-node" : "resource_23592",
    "meta" : {
      "weight" : 0.9,
      "name" : "isAuthor"
    }
  } ]
}

injections

In order to propagate (mis)credibility scores, you first need to tell us what are your initial ratings. This is done by the injections file. An example is provided:

{
  "injection" : {
    "resource_23586" : 1.0,
    "resource_23594" : 1.0
  }
}

This is a json file with a single object with property injection. The injection should be another object with as properties, the node identifiers used in the corresponding graph. The values are the initial (mis)credibility scores, they must be float values. They should be larger than 0.0. We recommend using values between 0.0 and 1.0. Ideally these will have been assigned by human raters.

Current limitations and improvement ideas

• It would be nice to extend the appleseed algorithm to accept a spreading-factor value that depends on the node type. This would allow us to control that some intermediate nodes should not keep (much of the) score, while other nodes will tend to keep most of the scores.
• When using the CLI, the client needs to specify the propagation weights when providing the graph. We could expose a way to assign default weights based on relationship types, but this would require providing additional information as part of the graph. Function kg-disinfo.generator.generate-test-graph already implements similar functionality which could be exposed and used for this. Similarly, the ability to automatically add inverse-edges could be specified as part of the graph metadata.

Bugs

• Not thoroughly tested

Acknowledgements

Originally developed as part of the DANTE project and refined as part of the Co-inform project.

License

Eclipse Public License 1.0 Copyright © 2019 Expert System Iberia