MultiStream Regression

Regression over a stream of data is challenging due to unbounded data size and non-stationary distribution over time. Typically, a traditional supervised regression model over a data stream is trained on data instances occurring within a short time period by assuming a stationary distribution. This model is later used to predict value of response-variable in future instances. Over time, the model may degrade in performance due to changes in data distribution among incoming data instances. Updating the model for change adaptation requires true value for every recent datainstances,which is scarce in practice. To overcome this issue, recent studies have employed techniques that sample fewer instances to be used for model re-training. Yet, this may introduce sampling bias that adversely affects the model performance. To address this scenario, we consider two independent, yet related, non-stationary data streams, which are referred to as the source and the target stream. The target stream continuously generates data instances whose value of response variable is unknown. The source stream, however, continuously generates data instances along with corresponding value for the response-variable, and has a biased data distribution with respect to the target stream. We refer to the problem of using a model trained on the biased source stream to predict the response-variable’s value in data instances occurring on the target stream as Multistream Regression. MSR offers a solution to the Multistream Regression problem. Please see the reference for more details.

MSR requires that-

• Input file will be provided in a ARFF/CSV format.
• All the features need to be numeric. If there is a non-numeric featues, those can be converted to numeric features using standard techniques.
• Features should be normalized to get better performance.

• Python 2.7
• Scipy, sklearn
• numpy, math

To execute the program:

1. First set properties in the config.properties file. Available options have been discussed later in this file.
2. Call the main function in the multistream.py file with two parameters. The first parameter is the path to the dataset file without extension. Extension is automatically appended from the corresponding property in the config.property file. The second parameter is the probability that the next instance will come from the source stream. As an example, the second parameter value 0.1 means that the next instance will come from the source stream with 10% probability and from the target stream with 90% probability.

• baseDir
  • Path to the base directory, which contains the input file(s). This will be appended to the name of the input file for getting the input file path.
• srcfileAppend
  • This string is appeneded after the name of input file supplied as the first parameter to get the file name for the source stream in the baseDir location.
• trgfileAppend
  • This string is appeneded after the name of input file supplied as the first parameter to get the file name for the target stream in the baseDir location.
• useKliepCVSigma
  • 1: Use the cross-validated value for sigma; 0: Use a fixed value for sigma.
• kliepDefSigma
  • In case useKliepCVSigma=0 was used, the value for sigma is specified in this property.
• kliepParEta
  • Value for the parameter Eta.
• kliepParLambda
  • Value for the parameter lambda.
• kliepParB
  • Value for the parameter B.
• kliepParThreshold
  • Value for the threshold used in the change detection algorithm.
• useSvmCVParams
  • If set, find the parameters for SVM using cross-validation.
• svmDefGamma
  • Default value for the gamma parameter in SVM.
• svmDefC
  • Default value for the parameter "C" in SVM.
• kernel
  • Type of kernel used in the svm algorithm.
• cushion
  • The value of cushion for the change detection algorithm if not calculated by gamma.
• sensitivity
  • Sensitivity of the change detection algorithm.
• maxWindowSize
  • Size of the source and target sliding window.
• initialDataSize
  • Size of the initial/warm-up training data.
• enableForceUpdate
  • If set, update the classifier after a long period of time even if there is no change detected.
• forceUpdatePeriod
  • If enableForceUpdate is set, the classifier is updated after this many instances even if there is no change detected.
• ensemble_size
  • Size of the ensemble.
• output_file_name
  • Path to the output file.
• logfile
  • Path to the log file.
• tempDir
  • Path to the directory containing all the temporary files.

• The program shows progress or any change point detected in console.
• At the end, it reports the overall accuracy.

1. A log file is generated in the location specified by "logfile" property, which contains important debug information.
2. The output file contains the running average accuracy.

A Framework for Multistream Regression With Direct Density Ratio Estimation