LodBrok is an LSTM model trained to detect the editor spam accounts in MusicBrainz. The model trained so far is done through offline batch training on the private editor dataset and will be integrated with MusicBrainz project through SpamNinja feature in the future.
Lodbrok receives four different inputs which are sub-arrays of the pre-processed input datum. The website and email inputs have respectively been tokenized to their top 1024 entries and are embedded into 256-dimensional vectors. Meanwhile, the user biography input is just reshaped into one 512-dimensional vector, as it is already quasi-embedded.
All three inputs are then passed into LSTMs where the bio-LSTM has an output twice as large as the others.
The outputs of the LSTMS are then concatenated with the other inputs (area set, non-zero privs, bio length, etc.) and passed into a stack of two fully-connected layers with 64 neurons and 50% dropout each.
The output layer consists of two neurons that represent the classification confidence for each category (spam and non-spam) and are activated using softmax so that their sum will always be one.
First, run the below command in a virtual environment(venv) with the terminal active in the root directory.
Install Python 3 requirements.
pip3 install -r requirements.txt
How to run each file in every folder is covered in the respective README.md.
Note: For handling datasets, please create a different virtual environement, go to data/README.md for information.
More details regarding datasets can be found here
More details regarding model training can be found here
More details regarding evaluation and prediction can be found here
In summary, the Lodbrok model achieves a very high spam detection rate while simultaneously maintaining a low false-positive rate. Data falsely classified by the model should be further examined to determine whether it really is part of the right dataset or whether there is a deficit of a certain type of data.