zhiyzuo / S3D

C++ code for rapid implementation of the Structured Sum-of-Squared Decomposition algorithm

structured sum of squares decomposition (s3d) algorithm

c++ code for fast implementation of the structured sum of squares decomposition (s3d) algorithm

note: this is a python wrapper. see original repo: https://github.com/peterfennell/S3D for pure c++ implementation

---

clarification

the design of the constructor is meant to be easy for hyperparameter tuning by cross validation. therefore, the methods fit and predict may be a little weird to use as standalone functions.

see a quickstart notebook for more

python dependencies

• pandas
• joblib
• networkx (the latest version has the edge arrow fixed!)
• matplotlib
• scikit-learn
• seaborn
• palettable

these can be installed by:

pip install -r requirements.txt

compiling s3d

compiling is simple and straight forward, do:

make all

or

make clean && make all

to remove previous compiled files.

data for cross val

for cross validation processes, datasets should be splitted first. the script split_data.py can be used to do so. the script can be run as the following:

python split_data.py data_name num_folds

this will read in data_name.csv from data/ folder; split data and store the train/test sets into splitted_data/data_name/ folder. for example, if there is a dataset called breastcancer.csv in data/, we can do:

python split_data.py breastcancer 5

if we do ls splitted_data/breastcancer/, we will see:

0  1  2  3  4

which are fold indices named folder, each of which there will be:

num_rows.csv  test.csv  train.csv

train.csv is the training set; test.csv is the testing set; num_rows.csv store the number of rows in train/test respectively.

finally, you can parallelize the data partitioning with more cores:

ptyhon split_data.py data_name num_folds num_jobs

where num_jobs is 1 by default.

data format

data format: in data_name.csv, the first column is the target column named target, followed by features:

target,feature_1,feature_2,...,feature_p

file structure

file directory setup: PYS3D class will by default create subfodlers data_name in:

tmp/ predictions/ models/ cv/

where:

• tmp will store the "inner cross validation" temporary files
• predictions will store the prediction results
• models will store the trained models
• cv will store the cross validation performance (on the validation set)

therefore, the first step is to create all these 4 for your convenience. you can do:

./init

which will cleanup and create these folders

reversely, run ./cleanup to remove these folders