This repository is the 2nd place solutions for the Kaggle - Open Problems - Multimodal Single-Cell Competition.
It contains two parts from senkin13 and tmp .
If you run into any trouble with the setup/code or have any questions please contact tmp at baosenguo@163.com and senkin13 at senkin13@hotmail.com.
This pipeline mainly consists of the following parts:
- Preprocessing
- FE
- Modeling
This simple solution produced a quite robust result (Public_lb 1st; Private_lb 2nd).
-
using raw count:
-
normalization:
-
transformation:
-
standardization:
-
batch-effect correction:
-
decomposition
- pca (64)
- ipca (128)
- factor analysis (64)
-
features selection
- Features highly correlated with target are selected. 245 features are selected in total.
-
cell-type (one-hot)
both mlp and lgb used the same features introduced above.
-
mlp (simple mlp performs best (single model with 1 seed - public 0.815; private 0.772))
-
lgb
- random 5-fold cv
- split according to "day"
-
dataset preparation
- /tmp/data/prepare.ipynb
- /tmp/data/preprocess.ipynb
-
training
- /tmp/model/lgb.ipynb
- /tmp/model/mlp.ipynb
- /tmp/model/blending.ipynb
- python 3.7.5
- pandas 1.3.5
- numpy 1.20.3
- torch 1.9.0
- sklearn 1.0.2
(The following specs were used to create the original solution)
Windows 10 (4 TB boot disk, 64 vCPUs, 300 GB memory) 1 x NVIDIA TITAN RTX
(python packages are detailed separately in requirements.txt):
Python 3.8.10 CUDA 11.3 cuddn 7.6.5.32 nvidia drivers v.466.47
(assumes the Kaggle API is installed)
below are the shell commands used in each step, as run from the top level directory
mkdir -p input features model sub
- download all data to input
DATA PROCESSING
- preprocess_cite.ipynb
- preprocess_multi.ipynb
TRAIN & PREDICTION
- cite_lgb_transformed_sparse_matrix.ipynb
- cite_lgb_raw_clr_pca.ipynb
- cite_lgb_raw_sparse_matrix.ipynb
- cite_lgb_raw_target.ipynb
- multi_lgb.ipynb
- multi_nn.ipynb
ENSEMBLE
- move tmp's tmp_cite_ensemble.joblib to ensemble/
- ensemble.ipynb