thu-syh / cikm-2022-analyticup-competition

For CIKM 2022 AnalytiCup Competition

Our code was explored and modified on the CIKM competition branch of the FS framework.

• To run the code:

  Considering the large difference in the number of data samples for each client, we follow the principle of non-leakage of data, refer to the data preprocessing process of usual model training, and perform oversampling and downsampling of the original data locally on each client, and adjust the ratio of training set to validation set.

    python resample_dataset.py

    python federatedscope/main.py --cfg federatedscope/gfl/baseline/isolated_cikmcup_impratio.yaml --client_cfg federatedscope/gfl/baseline/cikmcup_per_client_normalize.yaml data.root ../cikmdata_v2

• Brief introduction of the developed algorithm:

  We have tried a variety of federated learning algorithms, but unfortunately these methods have not shown good results. At present, our best result on the leaderboard is island training, but compared with the official baseline, certain configuration adjustments have been made. Such as learning rate of each client, Batchsize, LocalEpoch, etc. In addition, for some regression tasks, we counted the mean and standard deviation of y on the training set, and performed corresponding normalization and restoration operations during the training and prediction process.