In this project, we implemented several baselines (LightFM, DSSM, and SVD) and DeepFM model. We trained our models on the KION dataset.
We performed some preliminary preprocesing on the mentioned dataset. The notebook data_preprocessing.ipynb was used to procecss the data. After running this notebook, there should be 3 created csv files for items, users, and interactions respectively. These steps are enough to run our baselines (LightFM, DSSM, SVD)
For the data prepoccessing for DeepFM, the function preproc_deep_fm in utils/preprocessing.py were used. After running this function, it will combine the 3 csv files above to create 1 big csv. Alternatively, the preproccessed data can be downloaded here.
You can reproduce the results for the baselines in their respective notebooks (LIGHFM.ipynb, SVD.ipynb, and DSSR.ipynb). To reproduce the results for DeepFM, you can run the main.py script in src/DeepFM/main.py:
python3 main.py --train True --cold_start True
The train flag is to either train or validate. The cold_start flag is to either validate on cold start scenerio or standard scenerio.
Guo H., Tang R., Ye Y., Li Z. DeepFM: An End-to-End Wide & Deep Learning Framework for CTR Prediction. 2015