This is the code and dataset repository for UNICORN paper in ICMLCN 2025.

The traces that are used for training/test the models are included in the trace folder in the git repository. Full dataset before and after Colosseum can be found here.

The repository has the following overall structure:

• code directory: Contains the python code to do preprocessing, training, and test pipelines.
• traces directory: Contains all the KPI traces used in the paper UNICORN.
• file2app_map.pkl: Contains the mapping between the file names in traces and the application classes.
• Plotting.ipynb: Contains scripts to read prediction pickle files and plot the results.

As the first step you need to preprocess the dataset using preprocessing.py in the code directory. Run it as: ''' python preprocessing.py ''' The above script reads the KPI traces stored in the .csv files in the trace directory, preprocesses them and saves training, validation, and test partitions and statistics in the .pkl files stored in pkls directory.

Training and test pipeline can be run only after Preprocessing step is done.

The pipeline can be run through run_pipeline.sh in the code directory with the following content.

python top.py \
--gpu_id $1 \
--Train \
--Test \
--OOD_class 'Teams' \
--model_flag 'ResConvNet' \
--weight_filename 'weights-ICMLCN4-ResConvNet-5ids-oodindex4.pt'

This bash file runs the file top.py that calls the pipeline for training and test. Here is a short description of the arguments.

• --gpu_id: The ID of GPU node on your system. This can be set to 0 or 1 or 2 or ... depending on the GPU node you want to use.
• --Train: store_true parameter which indicates if we want to train an NN model.
• --Test: store_true parameter which indicates if we want to test a trained NN model.
• --OOD_class: The class name you would like to exclude from the training set. Can be any of the following: 'CallofDuty', 'Facebook', 'Meet', 'Zoom', 'Teams', 'Twitch'
• --model_flag: Indicating the model architecture. Can be 'ResConvNet' or ForwConvNet
• --weight_filename: Indicates the filename of NN weight files, that is used only for testing the trained model.

The pipeline can be run through:

./run_pipeline.sh 1

Where 1 is the GPU ID.

The test pipeline saves model predictions in .pkl files in the results directory.

You can plot the confusion matrices, bar graphs, line graphs, and scatter plots in the paper using Plotting.ipynb script.

My results directory for the paper UNICORN is located here, which can be downloaded and placed in the root folder next to code directory.