Reference

ArXiv -> ATL: Autonomous Knowledge Transfer from Many Streaming Processes

ResearchGate -> ATL: Autonomous Knowledge Transfer from Many Streaming Processes

Notes

If you want to see the original code used for this paper, access ATL_Matlab

ATL_Python is a reconstruction of ATL_Matlab by the same author, but using Python 3.6 and PyTorch (with autograd enabled and GPU support). The code is still not one-to-one and some differences in results can be found (specially on the data split methods in DataManipulator, however the network structure is correct and can be used by whoever is interested on this work in order to understand the structure or to build comparative results with your own research work.

Having said that, expect ATL_Python to be updated in the following weeks, including functions refactoring and functions documentation.

ATL_Python

ATL: Autonomous Knowledge Transfer From Many Streaming Processes ACM CIKM 2019

1. Clone ATL_Python git to your computer, or just download the files.

2. Provide a dataset by replacing the file data.csv The current data.csv holds https://www.researchgate.net/publication/221653408_A_Streaming_Ensemble_Algorithm_SEA_for_Large-Scale_Classification dataset. data.csv must be prepared as following:

- Each row presents a new data sample
- Each column presents a data feature
- The last column presents the label for that sample. Don't use one-hot encoding. Use a format from 1 onwards

3. Open Matlab. The code was developed using Matlab 2018b, so if you use an older version, you might get some incompability errors.

You can use Matlab 2018b or newer. Matlab may prompt you to install some official add-ons, as:

- Deep Learning Toolbox
- Fuzzy Logic Toolbox
- Digital Processing Signal Toolbox

4. Inside Matlab, travel until the folder where you downloaded ATL_Matlab.

5. On the Matlab terminal, just type ATL. This will execute ATL, which will read your data.csv and process it.

ATL will automatically normalize your data and split your data into 2 streams (Source and Target data streams) with a bias between them, as described in the paper.

Matlab will print ATL status at the end of every minibatch, where you will be able to follow useful information as:

- Training time (maximum, mean, minimum, current and accumulated)
- Testing time (maximum, mean, minimum, current and accumulated)
- The number of GMM clusters (maximum, mean, minimum and current)
- The target classification rate
- And a quick review of ATL structure (both discriminative and generative phases), where you can see how many automatically generated nodes were created.

At the end of the process, Matlab will plot 6 graphs:

- Network bias and Network variance w.r.t. the generative phase
- Network bias and Network variance w.r.t. the discriminative phase
- The target and source classification rate evolution, as well as the final mean accuracy of the network
- All losses over time, and how they influence the network learning
- The evolution of GMMs on Source and Taret AGMMs over time
- The processing time per mini-batch and the total processing time as well, both for training and testing

Thank you.

Download all datasets used on the paper

As some datasets are too big, we can't upload them to GitHub. GitHub has a size limite of 35MB per file. Because of that, you can find all the datasets in a csv format on the anonymous link below. To test it, copy the desired dataset to the same foler as ATL and rename it to data.csv.

• https://drive.google.com/open?id=1emgVw6muSodzozQcuz7ks8XeZYPxGEZ7