Class StreamingLinearAlgorithm<M extends GeneralizedLinearModel,A extends GeneralizedLinearAlgorithm<M>>

Object
org.apache.spark.mllib.regression.StreamingLinearAlgorithm<M,A>
All Implemented Interfaces:
org.apache.spark.internal.Logging
Direct Known Subclasses:
StreamingLinearRegressionWithSGD, StreamingLogisticRegressionWithSGD

public abstract class StreamingLinearAlgorithm<M extends GeneralizedLinearModel,A extends GeneralizedLinearAlgorithm<M>> extends Object implements org.apache.spark.internal.Logging
StreamingLinearAlgorithm implements methods for continuously training a generalized linear model on streaming data, and using it for prediction on (possibly different) streaming data.

This class takes as type parameters a GeneralizedLinearModel, and a GeneralizedLinearAlgorithm, making it easy to extend to construct streaming versions of any analyses using GLMs. Initial weights must be set before calling trainOn or predictOn. Only weights will be updated, not an intercept. If the model needs an intercept, it should be manually appended to the input data.

For example usage, see StreamingLinearRegressionWithSGD.

NOTE: In some use cases, the order in which trainOn and predictOn are called in an application will affect the results. When called on the same DStream, if trainOn is called before predictOn, when new data arrive the model will update and the prediction will be based on the new model. Whereas if predictOn is called first, the prediction will use the model from the previous update.

NOTE: It is ok to call predictOn repeatedly on multiple streams; this will generate predictions for each one all using the current model. It is also ok to call trainOn on different streams; this will update the model using each of the different sources, in sequence.

  • Constructor Details

    • StreamingLinearAlgorithm

      public StreamingLinearAlgorithm()
  • Method Details

    • latestModel

      public M latestModel()
      Return the latest model.

      Returns:
      (undocumented)
    • predictOn

      public DStream<Object> predictOn(DStream<Vector> data)
      Use the model to make predictions on batches of data from a DStream

      Parameters:
      data - DStream containing feature vectors
      Returns:
      DStream containing predictions

    • predictOn

      public JavaDStream<Double> predictOn(JavaDStream<Vector> data)
      Java-friendly version of predictOn.

      Parameters:
      data - (undocumented)
      Returns:
      (undocumented)
    • predictOnValues

      public <K> DStream<scala.Tuple2<K,Object>> predictOnValues(DStream<scala.Tuple2<K,Vector>> data, scala.reflect.ClassTag<K> evidence$1)
      Use the model to make predictions on the values of a DStream and carry over its keys.
      Parameters:
      data - DStream containing feature vectors
      evidence$1 - (undocumented)
      Returns:
      DStream containing the input keys and the predictions as values

    • predictOnValues

      public <K> JavaPairDStream<K,Double> predictOnValues(JavaPairDStream<K,Vector> data)
      Java-friendly version of predictOnValues.

      Parameters:
      data - (undocumented)
      Returns:
      (undocumented)
    • trainOn

      public void trainOn(DStream<LabeledPoint> data)
      Update the model by training on batches of data from a DStream. This operation registers a DStream for training the model, and updates the model based on every subsequent batch of data from the stream.

      Parameters:
      data - DStream containing labeled data
    • trainOn

      public void trainOn(JavaDStream<LabeledPoint> data)
      Java-friendly version of trainOn.
      Parameters:
      data - (undocumented)