Gaussian Receptive Fields for Stream Learning
Stream data processing has gained progressive momentum with the arriving of new stream applications and big data scenarios. One of the most promising Spiking Neural Network technique for stream learning is the Evolving Spiking Neural Network, which uses an interesting population encoding scheme (the Gaussian receptive fields) to transform the incoming stimuli into spikes. These files have been used to carried out a study that sheds lights on their use for stream learning, and focuses on applying them as a pre-processing technique to any dataset in order to gain representativeness, and to boost the predictive performance of the stream learning methods. Experiments with synthetic and real data sets have been carried out, and lead to think that their use could be applied successfully as a general pre-processing technique in many real cases.
- Required frameworks: scikit-multiflow (https://scikit-multiflow.github.io/) and scikit-learn
- Dependencies: texttable, collections, statsmodels.stats.contingency_tables, pandas, numpy, math, warnings, pickle, scipy.io, scipy.stats, matplotlib.pyplot, seaborn
The "benchmark_git.py" script is used to generate the final results. It uses the scikit-learn framework (GaussianNB, PassiveAgressiveClassifier, MLPClassifier, and SGDClassifier techniques). It also uses the scikit-multiflow framework (HoeffdingTree, HoeffdingAdaptiveTree, and KNN techniques, and also the ADWIN drift detector).
The "evaluate_prequential_NN.py" script is used for the streaming evaluation. This file should be placed in the corresponding folder of the scikit-multiflow package: '.../scikit-multiflow-master/src/skmultiflow/evaluation'
The following scripts are modified versions of the original algorithms for the scikit-multiflow framework in order to consider the GRF attributes. They should be placed in the corresponding folder. They are:
- GRF_HoeffdingTree.py--> '.../scikit-multiflow-master/src/skmultiflow/trees/'
- GRF_HoeffdingAdaptiveTree.py--> '.../scikit-multiflow-master/src/skmultiflow/trees/'
- GRF_KNN.py--> '.../scikit-multiflow-master/src/skmultiflow/lazy/'
The following scripts are modified versions of the original algorithms for the scikit-learn framework in order to consider the GRF attributes. They should be placed in the corresponding folder. They are:
- GRF_GaussianNB.py--> '.../lib/python3.6/site-packages/sklearn'
- GRF_PassiveAgressiveClassifier.py--> '.../lib/python3.6/site-packages/sklearn/linear_model'
- GRF_MLPClassifier.py--> '.../lib/python3.6/site-packages/sklearn/neural_network'
- GRF_SGDClassifier.py--> '.../lib/python3.6/site-packages/sklearn/linear_model'
The following synthetic datasets are provided:
- circle1 (first concept of the circleG dataset)
- circle2 (second concept of the circleG dataset)
- line1 (first concept of the line dataset)
- line2 (second concept of the line dataset)
- sine1 (first concept of the sineV dataset)
- sine2 (second concept of the sineV dataset)
- sineH1 (first concept of the sineH dataset)
- sineH2 (second concept of the sineH dataset)
- SEA
The following real-world datasets are provided:
- weather
- electricity
- moving squares
- Airlines