Dynamic Weighted Ensemble - Local Fusion

This repository contains an implementation for the Dynamic Weighted Ensemble (DWE) - Local Fusion method. Find the paper in this ref on IEEE.

Local Fusion is an ensemble techinque that could be used to improve predictions by weighing appropriately the single models contribution.

Installation

Pypi

	pip install ensemblem

Docs - ReadTheDocs (UnderReview)

Usage

First of all, you need to define the KWEnsembler class. Next, it's required to provide the search-space (it could be the validation set / neighbours-set) in which the ensembler will find the nearest elements to the generic test sample.

	from ensemblem.model import KWEnsembler
	ensemble = KWEnsembler(5)
	ensemble.fit(X_validation, y_validation)

Finally, calling the prediction method the class will produce the forecasts.

	ensemble.predict(X_test,
                    features_space,
                    other_model_prediction_columns)

The class returns predictions in the same order in which they are provided. It supports one or multiple samples to forecasts. In this library, we refers to the neighbours-set as the space in which the ensembler will find the nearest elements to the generic test sample.

Example of using the KWEnsembler class

You can find a detailed tutorial in the readthedocs webpage.

1. Load data
2. Split data into train, neighbours-set and test sets
3. Train multiple expert models on the train data
4. Generate predictions for the test data
5. Train the ensembler on neighbours-set
6. Generate predictions for the test dataset using ensembler
7. Compare the predictions from the ensembler with the predictions from the expert models

Results & Benchmarks

Credits

Algorithm Applications

• A dynamic weighting ensemble approach for wind energy production prediction IEEE

• An ensemble approach to sensor fault detection and signal reconstruction for nuclear system control Elsevier

Possible Improvements

• [Docs] General improvements over documentations

• [Code] Clean-code

• [Engineering] When dealing with features coming with magnitude and different meaning, it's relevant to normalize values appropriately.

• [Engineering] Additional measures to the simple euclidean-space

Licence

The code is provided with a MIT licence.