confStream: Automated Algorithm Configuration for Stream Clustering

[MatthiasCarnein]

Hi, this pull request adds our confStream algorithm to MOA. confStream allows to automatically find the best stream clustering algorithm and its best parameter configuration for a given data stream in real-time. As a result, the user no longer has to choose certain parameter settings but the algorithm can do this automatically and over time. In our results, this has improved the clustering quality considerably.

For this, the algorithm trains an ensemble of different algorithms and configurations in parallel. Periodically, the most promising configurations will create offsprings which can improve the overall quality over time. A brief overview of this is shown below and more information is available in the corresponding papers:

The initial idea has been presented at the Workshop on Automated Data Science at ECML PKDD '19. More detailed results will be presented at this years LION 14 conference.

• Carnein M, Trautmann H, Bifet A and Pfahringer B (2020), "confStream: Automated Algorithm Selection and Configuration of Stream Clustering Algorithms", To appear in 14th Learning and Intelligent Optimization Conference (LION 14)
• Carnein M, Trautmann H, Bifet A and Pfahringer B (2019), "Towards Automated Configuration of Stream Clustering Algorithms", In Workshop on Automated Data Science at the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD '19)

An overview of the algorithm and the published papers is also available here: https://www.carnein.com/confstream

The general idea of confStream is inspired by the BLAST algorithm:

• van Rijn JN, Holmes G, Pfahringer B and Vanschoren J (2015), "Having a Blast: Meta-Learning and Heterogeneous Ensembles for Data Streams", In Proceedings of the 2015 IEEE International Conference on Data Mining (ICDM '15)

The algorithm can be selected like any other clustering algorithm in MOA. However, the algorithm makes use of a configuration file which lists all the algorithms and their parameters to optimise. All stream clustering algorithms in MOA as well as all their parameters can be selected.

An example configuration file is available in "moa/moa/src/main/java/moa/clusterers/meta/" and also shown below. Most importantly, the configuration file specifies a list of algorithms and their parameters to tune. Here, DenStream and ClusTree as well as all their relevant parameters are tuned.

{
    "windowSize": 1000,
    "ensembleSize": 20,
    "newConfigurations": 10,
    "keepCurrentModel": "true",
    "reinitialiseWithClusters": "true",
    "preventAlgorithmDeath": "true",
    "evaluateMacro": "false",
    "keepGlobalIncumbent": "true",
    "keepAlgorithmIncumbents": "true",
    "keepInitialConfigurations": "true",
    "useTestEnsemble": "true",
    "lambda": 0.05,
    "resetProbability": 0.01,
    "numberOfCores": 1,
    "performanceMeasure": "SilhouetteCoefficient",
    "performanceMeasureMaximisation": "true",
    "algorithms": [
        {
            "algorithm": "denstream.WithDBSCAN",
            "parameters": [
                {"parameter": "e", "type":"numeric", "value":0.02, "range":[0,1]},
                {"parameter": "b", "type":"numeric", "value":0.2, "range":[0,1]},
                {"parameter": "m", "type":"integer", "value":1, "range":[0,10000]},
                {"parameter": "o", "type":"integer", "value":2, "range":[2,20]},
                {"parameter": "l", "type":"numeric", "value":0.25, "range":[0,1]}
            ]
        }
        ,
        {
            "algorithm": "clustree.ClusTree",
            "parameters": [
                {"parameter": "H", "type":"integer", "value":8, "range":[1,20]},
                {"parameter": "B", "type":"boolean", "value":"false"}
            ]
        }
    ]
}