This repository contains the code used for the experimentation shown in the paper that will be presented at IEEE International Conference on Big Data 2024.
execute:
conda create -n env python=3.8
conda activate env
pip install -r requirements.txt
The project is composed of the following directories.
Download datasets.zip from here.
Extract it in the main folder of the project.
It contains the generated data streams. Each file's name has the following structure: <generator>_federated_<id_configuration>conf_node<id_device>_<id_concept>task_<start>.csv.
Generators:
- sine: SRW.
- weather: Weather
- air_quality: AirQuality.
For each configuration, the data stream's concepts of each device are split into different csv files.
Concept IDs are incremental and starts from 0.
In case of devices 1 and 2, the first concept's data stream (task0) is split in two parts to start the following device at the end of the first concept's initial part. The initial part is indicated as "task0_start".
It contains the python modules implementing cPNN, cLSTM and MAcPNN.
It contains the python modules to implement the prequential evaluation used for the experiments.
It contains the python modules implementing the data stream generator.
It runs the prequential evaluation using the specified configurations. Change the variables in the code for different settings (see the code's comments for the details).
Run it with the command python -m evaluation.test_ma.
The execution stores the pickle files containing the results in the folder specified by the variable PATH_PERFORMANCE. For the details about the pickle files, see the documentation in evaluation/prequential_evaluation.py.
After executing the experiments with test_ma.py, it builds the performance table and write it as csv and xlsx.
The column T<i>_<metric>_<model> represents the metric (start, end) associated with concept i.
The column <metric>_<model> represents the metric (start, end) averaged on the concepts (the first concept is excluded).
Run it with the command python -m evaluation.build_perf_table.
After executing the experiments with test_ma.py, it builds the performance files to make the plots in Fig. 7.
Run it with the command python -m evaluation.build_perf_plot.
After making the files with build_perf_plot.py, it makes the plots in Fig. 7 and saves it in the performance/macpnn folder.
Run it with the command python -m evaluation.plot.
https://github.com/AndreaCossu/ContinualLearning-SequentialProcessing