StreamLab: Large-Scale Machine Learning on Streams
Sameer Singh, Ignacio Cano, Tianqi Chen, Marco Ribeiro, Carlos Guestrin

Citation
Sameer Singh, Ignacio Cano, Tianqi Chen, Marco Ribeiro, Carlos Guestrin. "StreamLab: Large-Scale Machine Learning on Streams". Talk or presentation, 28, October, 2014; Poster presented at the 2014 TerraSwarm Annual Meeting. .

Abstract
We are exploring a novel abstraction for distributed predictive modeling over streaming data, called StreamLab. Existing architectures, such as Graphlab, are designed to handle large collections of structured data, however the same modeling paradigms do not hold when data is in form of multiple streams of high-volume items. The team is exploring machine learning for parallel, online training that can capture dependencies across streams, and designing primitives that can be easily used to implement a large family of complex learning algorithms. Along with the implementation that includes a distributed parameter server that allows parallel online learners to share parameters, and stream aggregators that compute approximate statistics over large streams, the team has developed a novel “learning over counts” algorithm. This training algorithm uses aggregated statistics for prediction, and experiments on a variety of datasets from KDDCup and Kaggle demonstrate that “learning by counts” provides a scalable, accurate, and generalizable machine learning approach. Future work on StreamLab includes design and implementation of a number of existing machine learning algorithms into the current framework, in the process identifying primitive operations and reusable modules common to large-scale stream analytics.

Electronic downloads

Internal. This publication has been marked by the author for TerraSwarm-only distribution, so electronic downloads are not available without logging in.

Citation formats

• HTML

  Sameer Singh, Ignacio Cano, Tianqi Chen, Marco Ribeiro,
  Carlos Guestrin. <a
  href="http://www.terraswarm.org/pubs/455.html"><i>StreamLab:
  Large-Scale Machine Learning on Streams</i></a>,
  Talk or presentation,  28, October, 2014; Poster presented
  at the <a
  href="http://www.terraswarm.org/conferences/14/annual"
  >2014 TerraSwarm Annual Meeting</a>.
  .

• Plain text

  Sameer Singh, Ignacio Cano, Tianqi Chen, Marco Ribeiro,
  Carlos Guestrin. "StreamLab: Large-Scale Machine
  Learning on Streams". Talk or presentation,  28,
  October, 2014; Poster presented at the <a
  href="http://www.terraswarm.org/conferences/14/annual"
  >2014 TerraSwarm Annual Meeting</a>.
  .

• BibTeX

  @presentation{SinghCanoChenRibeiroGuestrin14_StreamLabLargeScaleMachineLearningOnStreams,
      author = {Sameer Singh and Ignacio Cano and Tianqi Chen and
                Marco Ribeiro and Carlos Guestrin},
      title = {StreamLab: Large-Scale Machine Learning on Streams},
      day = {28},
      month = {October},
      year = {2014},
      note = {Poster presented at the <a
                href="http://www.terraswarm.org/conferences/14/annual"
                >2014 TerraSwarm Annual Meeting</a>.
  },
      abstract = {We are exploring a novel abstraction for
                distributed predictive modeling over streaming
                data, called StreamLab. Existing architectures,
                such as Graphlab, are designed to handle large
                collections of structured data, however the same
                modeling paradigms do not hold when data is in
                form of multiple streams of high-volume items. The
                team is exploring machine learning for parallel,
                online training that can capture dependencies
                across streams, and designing primitives that can
                be easily used to implement a large family of
                complex learning algorithms. Along with the
                implementation that includes a distributed
                parameter server that allows parallel online
                learners to share parameters, and stream
                aggregators that compute approximate statistics
                over large streams, the team has developed a novel
                �learning over counts� algorithm. This
                training algorithm uses aggregated statistics for
                prediction, and experiments on a variety of
                datasets from KDDCup and Kaggle demonstrate that
                �learning by counts� provides a scalable,
                accurate, and generalizable machine learning
                approach. Future work on StreamLab includes design
                and implementation of a number of existing machine
                learning algorithms into the current framework, in
                the process identifying primitive operations and
                reusable modules common to large-scale stream
                analytics.},
      URL = {http://terraswarm.org/pubs/455.html}
  }
  

Posted by Sameer Singh on 10 Nov 2014.
Groups: tools