-
Notifications
You must be signed in to change notification settings - Fork 328
7. Clustering of Submissions
By default, JPlag is configured to perform a clustering of the submissions. The clustering partitions the set of submissions into groups of similar submissions. The found clusters can be used candidates for potentially colluding groups. Each cluster has a strength score, that measures how suspicious the cluster is compared to other clusters.
Clustering can take a long when there is a large number of submissions.
Users who are not interested in clustering can safely disable it with the --cluster-skip
option.
Clustering can either be configured using the CLI options or programmatically using the ClusteringOptions
class. Both options work analogously and share the same default values.
The clustering is designed to work out-of-the-box for running within the magnitude of about 50-500 submissions, but it can be tweaked when problems occur. For more submissions it might be necessary to increase Max-Runs
or Bandwidth
, so that an appropriate number of clusters can be determined.
Group | Option | Description | Default |
---|---|---|---|
General | Enable | Controls whether the clustering is run at all. | true |
General | Algorithm | Which clustering algorithm to use.
|
Agglomerative Clustering |
General | Metric | The similarity score between submissions to use during clustering. Each score is expressed in terms of the size of the submissions A and B and the size of their matched intersection A ∩ B .
|
AVG |
Spectral | Bandwidth | For Spectral Clustering, Baysian Optimization is used to determine a fitting number of clusters. If a good clustering result is found during the search, numbers of clusters that differ by something in range of the bandwidth are also expected to good. Low values result in more exploration of the search space, high values in more exploitation of known results. | 20.0 |
Spectral | Noise | The result of each k-Means run in the search for good clusterings is random. The noise level models the variance in the "worth" of these results. It also acts as a regularization constant. | 0.0025 |
Spectral | Min-Runs | Minimum number of k-Means executions for spectral clustering. With these initial runs clustering sizes are explored. | 5 |
Spectral | Max-Runs | Maximum number of k-Means executions during spectral clustering. Any execution after the initial (min-) runs tries to balance between exploration of unknown clustering sizes and exploitation of clustering sizes known as good. | 50 |
Spectral | K-Means Iterations | Maximum number of iterations during each execution of the k-Means algorithm. | 200 |
Agglomerative | Threshold | Only clusters with an inter-cluster-similarity greater than this threshold are merged during agglomerative clustering. | 0.2 |
Agglomerative | inter-cluster-similarity | How to measure the similarity of two clusters during agglomerative clustering.
|
AVERAGE |
Preprocessing | Pre-Processor | How the similarities are preprocessed prior to clustering. Spectral Clustering will probably not have good results without it.
|
CDF |
All clustering related classes are contained within the de.jplag.clustering(.*)
packages in the core project.
The central idea behind the structure of clustering is the ease of use: To use the clustering calling code should only ever interact with the ClusteringOptions
, ClusteringFactory
, and ClusteringResult
classes:
classDiagram
ClusteringFactory <.. CallingCode
ClusteringOptions <.. CallingCode : creates
ClusteringAdapter <.. ClusteringFactory
ClusteringAlgorithm <.. ClusteringAdapter : runs
ClusteringAlgorithm <.. ClusteringFactory : creates instances
ClusteringPreprocessor <.. ClusteringFactory : creates instances
PreprocessedClusteringAlgorithm <.. ClusteringFactory : creates
ClusteringOptions <-- ClusteringFactory
ClusteringAlgorithm <|-- PreprocessedClusteringAlgorithm
ClusteringAlgorithm <-- PreprocessedClusteringAlgorithm : delegates to
ClusteringPreprocessor ..o PreprocessedClusteringAlgorithm
class ClusteringFactory{
getClusterings(List~JPlagComparison~ comparisons, ClusteringOptions options)$ ClusteringResult~Submission~
}
class ClusteringOptions{
}
class ClusteringAlgorithm {
<<interface>>
cluster(Matrix similarities) ClusteringResult~Integer~
}
class ClusteringPreprocessor {
<<interface>>
preprocess(Matrix similarities) Matrix
}
class ClusteringAdapter{
ClusteringAdapter(List~JplagComparison~ comparisons)
doClustering(ClusteringAlgorithm algorithm) ClusteringResult~Submission~
}
class PreprocessedClusteringAlgorithm{
cluster(Matrix similarities) ClusteringResult~Integer~
}
class CallingCode{
}
New clustering algorithms and preprocessors can be implemented using the GenericClusteringAlgorithm
and ClusteringPreprocessor
interfaces which operate on similarity matrices only. ClusteringAdapter
handles the conversion between de.jplag
classes and matrices. PreprocessedClusteringAlgorithm
adds a preprocessor onto another ClusteringAlgorithm
.
- based on On Spectral Clustering: Analysis and an algorithm (Ng, Jordan & Weiss, 2001)
- automatic hyper-parameter search using Bayesian Optimization with a Gaussian Process as the surrogate model and L-BFGS for optimization on the surrogate
- the L-BFGS implementation is a pit of technical debt, see here.
There are integration tests for the Spectral Clustering to verify, that a least in the case of two known sets of similarities the groups known to be colluders are found. However, these are considered to be sensitive data. The datasets are not available to the public and these tests can only be run by maintainers with access.
To run these tests the contents of the PseudonymizedReports repository must added in the folder jplag/src/test/resources/de/jplag/PseudonymizedReports
.