miGraph / MIGraph

This repository reproduces the miGraph / MIGraph algorithms from the paper Multi-instance learning by treating instances as non-I.I.D. samples.

Only the core algorithm code is tracked. Data preprocessing scripts, timing scripts, IDE files, generated caches, and dataset files are intentionally left out of git.

Requirements

pip install -r requirements.txt

Run

The code expects MIL data in the usual MATLAB format, where the .mat file contains a data variable. Each row is a bag: data[i, 0] is the instance matrix and data[i, 1] is the bag label. Instance labels, when present, should be stored in the last column of each instance matrix.

Build the miGraph precomputed kernel folds:

python -c "from miGraph import miGraph; model = miGraph('path/to/data.mat'); train_K, train_y, test_K, test_y, _ = next(model.get_mapping()); print(train_K.shape, test_K.shape)"

Build the MIGraph precomputed kernel folds:

python -c "from MIGraph_explicit import MIGraph; model = MIGraph('path/to/data.mat'); train_K, train_y, test_K, test_y, _ = next(model.get_mapping()); print(train_K.shape, test_K.shape)"

The returned matrices are precomputed kernels and can be passed to a classifier that supports precomputed kernels.