gaohailong / moa-abfs

MOA-ABFS

Boosting Decision Stumps for Dynamic Feature Selection on Data Streams (ABFS) - INFORMATION SYSTEMS SUBMISSION

This repository includes the code for the ABFS feature selector. The code here reflects the code used in the submission entitled "Boosting Decision Stumps for Dynamic Feature Selection on Data Streams" authored by Jean Paul Barddal, Heitor Murilo Gomes, Fabrício Enembreck, Albert Bifet, and Bernhard Pfahringer. This paper has been submitted to Information Systems for refereeing.

In practice, this repository is a clone of the Massive Online Analysis (MOA) framework, with the addition of:

1. The code for the proposed method (BoostingSelector.java): this class represents the ABFS method proposed in the paper.
2. Abstract classes and interfaces for future feature selectors for data streams (AbstractFeatureSelector.java and FeatureSelector.java): these files contain the generic behavior that is shared between feature selectors. In practice, a feature selector should be able to be updated with labeled instances and return the currently selected subset of features at any time.
3. The gold-standard selector (GoldStandardSelector.java): Referred to as ORACLE in the paper, this selector selects the correct features for each instance if the generator has this information available.
4. The code for stability computation (EvaluateFeatureSelectionStability.java): This class is responsible for calculating the Stability of a feature selector on a data stream. An important parameter is the validation scheme to be adopted (bootstrap-, split- or cross-validation). Details about these schemes are available in the paper.
5. The code for Selection Accuracy computation (Part of the AbstractFeatureSelector.java file): to facilitate the computation of feature selection-specific metrics,
6. The scripts to reproduce the experiments discussed in the paper (ABFSExperiments.java and StabilityExperiments.java):
7. Synthetic data generators (BG.java, BG2.java, and BG3.java): These data generators have been introduced in one of our previous papers entitled "A survey on feature drift adaptation: Definition, benchmark, challenges and future directions" borrowing ideas from Mark Hall's thesis and were made available for MOA.

Upon paper acceptance, the authors plan to make the code available here part of the official repository of MOA (see details below).

More about MOA

If you want to know more about MOA, please see the following paper:

Albert Bifet, Geoff Holmes, Richard Kirkby, Bernhard Pfahringer (2010); MOA: Massive Online Analysis; Journal of Machine Learning Research 11: 1601-1604

Also, the official GitHub repository for MOA can be found here.