Lazy and Fast Greedy MAP Inference for Determinantal Point Process

This code is the official implementation of Lazy and Fast Greedy MAP Inference for Determinantal Point Process.

Requirements

• CMake (version ≧ 3.23)
• pkg-config
• GNU Make or Ninja
• C++ Compiler (GNU Compiler Collection (GCC) / Clang / ...) compatible to C++17
  • GCC: version 7.1 or later
  • Clang: version 5.0 or later

Compile

When first cloning this repository, run the following commands:

git submodule init
git submodule update

To compile C++ codes, run:

cmake --preset make  # replace "make" with "ninja" if you use Ninja
cmake --build --preset release

Data Preprocessing

To generate the input data used in the experiment, follow these steps. The resulting data will be stored to data/.

Synthetic Datasets

To generate synthetic data, run the following:

./build/gen_wishart

Real-world Datasets

To pre-process the real world datasets, Please follow these steps:

MovieLens 25M

To get the primary data of MovieLens 25M dataset, run the following commands:

mkdir -p data
wget -P data https://files.grouplens.org/datasets/movielens/ml-25m.zip
unzip data/ml-25m.zip -d data
./build/pre_process -d movie_lens

Netflix Prize

To get Netflix Prize dataset, you need a Kaggle account. Logging to Kaggle, download archive.zip from here and store it to data/. For pre-processing, run the following commands.

mkdir -p data
unzip data/archive.zip -d data/netflix_raw
./build/pre_process -d netflix

Computing Product Matrices

The matrix $L = B^\top B$ for Real-world datasets can be computed by the following (run on the root directory):

./build/product -d netflix
./build/product -d movie_lens

Run Experiments

Run commands on the root directory.

Greedy, RandomGreedy, StochasticGreedy, InterlaceGreedy

./build/exp -a [algorithm] -d [dataset_name] -m [input_matrix]

• algorithm: greedy (default), random, stochastic, interlace
• dataset_name: wishart, wishart_fixed_k, movie_lens, netflix
• input_matrix: B (default), L

DoubleGreedy

./build/double -d [dataset_name]

• dataset_name: wishart, movie_lens, netflix

Experimental results will be stored to result/ in the CSV format.

License

The code is licensed MIT.