Legendre Decomposition for Tensors

An implementation of Legendre decomposition for tensors, which decomposes a given nonnegative tensor into a multiplicative combination of parameters. Current implementation supports only third order tensors and zero or nonnegative values are directly ignored. Please see the following paper for more details:

• Sugiyama, M., Nakahara, H., Tsuda, K.: Legendre Decomposition for Tensors, NIPS 2018 (to appear).

Usage

In your program

You can perform tensor balancing by calling the function LegendreDecomposition. To use it, you just need to include the header file "legendre_decomposition.h" in your program. The code is written in C++11 and the Eigen library is needed.

The main function LegendreDecomposition is defined as:

double LegendreDecomposition(Tensor& X, Int core_size, double error_tol, double rep_max, bool verbose, int type, int const_type, int *num_param)

• X: an input tensor, the type Tensor is defined as vector<vector<vector<double>>>
• core_size: the parameter for a decomposition basis
• error_tol: error tolerance
• rep_max: the maximum number of iteration
• verbose: the verbose mode if true
• type: type of an algorithm
  • type == 1 Natural gradient (recommended)
  • type == 2 Gradient descent
• const_type: type of a decomposition basis (currently can be 1 [complex] or 2 [simple])
• num_param: the number of parameters will be returned
• Return value: the number of iterations

In terminal

We provide a sample tensor "test.csv" and a test code "main.cc" to try the code, which includes an input and output interface for tensor files.

For example, in the directory src/cc:

$ make
$ ./ld -d 4 -i test.csv -c 2
> Read a database file "test.csv":
  Size: (3, 5, 4)
        (Note: this is treated as (4, 3, 5) inside the implementation)
> Start Legendre decomposition by natural gradient:
  Number of parameters: 19
  Step  1, Residual: 0.00419722
  Step  2, Residual: 2.08951e-05
  Step  3, Residual: 5.15477e-10
> Profile:
  Number of iterations: 3
  Running time:         0.000205 [sec]
  RMSE:                 0.222512

To compile the program, please edit paths in the "Makefile" according to the location of the Eigen library in your environment.

Command-line arguments

• -i <input_file>: a path to a csv file of an input tensor (without row and column names)
• -o <output_matrix_file>: an output file of the reconstructed tensor
• -t <output_stat_file>: an output file of statistics
• -e <error_tolerance>: error tolerance is set to 1e-<error_tolerance> [default value: 5]
• -r <max_iteration>: the maximum number of iterations is set to 1e+<max_iteration> [default value: 6]
• -v: the verbose mode if specified
• -n: the natural gradient is used
• -g: the gradient descent is used
• -d: the depth size (the number of matrices)
• -c: the parameter for a decomposition basis
• -b: type of a decomposition basis [default value: 1]

Contact

Author: Mahito Sugiyama
Affiliation: National Institute of Informatics, Tokyo, Japan
E-mail: mahito@nii.ac.jp