bmcfee / mlr

Metric Learning to Rank
GNU General Public License v3.0
46 stars 39 forks source link

Metric Learning to Rank (mlr-1.2) http://www-cse.ucsd.edu/~bmcfee/code/mlr/

AUTHORS: Brian McFee bmcfee@cs.ucsd.edu Daryl Lim dklim@ucsd.edu

This code is distributed under the GNU GPL license. See LICENSE for details or http://www.gnu.org/licenses/gpl-3.0.txt

INTRODUCTION

This package contains the MATLAB code for Metric Learning to Rank (MLR).
The latest version of this software can be found at the URL above.

The software included here implements the algorithm described in

[1] McFee, Brian and Lanckriet, G.R.G. Metric learning to rank.
Proceedings of the 27th annual International Conference on Machine Learning (ICML), 2010.

Please cite this paper if you use this code.

If you use the Robust MLR code (rmlr_train), please cite:

[2] Lim, D.K.H., McFee, B. and Lanckriet, G.R.G. Robust structural metric learning. Proceedings of the 30th annual International Conference on Machine Learning (ICML), 2013.

INSTALLATION

  1. Requirements

This software requires MATLAB R2007a or later. Because it makes extensive use of the "bsxfun" function, earlier versions of Matlab will not work.

If you have an older Matlab installation, you can install an alternative bsxfun implementation by going to

http://www.mathworks.com/matlabcentral/fileexchange/23005 ,

however, this may not work, and it will certainly be slower than the native bsxfun implementation.

  1. Compiling MEX functions

MLR includes two auxiliary functions "cummax" and "binarysearch" to accelerate certain steps of the optimization procedure. The source code for these functions is located in the "util" subdirectory.

A makefile is provided, so that (on Unix systems), you can simply type

cd util make cd ..

  1. Running the demo

To test the installation, first add the path to MLR to your Matlab environment:

>> addpath(genpath('/path/to/mlr/'));

Then, run the demo script:

>> mlr_demo

from within Matlab. The demo generates a random training/test split of the Wine data set (http://archive.ics.uci.edu/ml/datasets/Wine) and learns a metric with MLR. Both native and learned metrics are displayed in a figure with a scatter plot.

TRAINING

There are several modes of operation for training metrics with MLR. In the simplest mode, training data is contained in a matrix X (each column is a training vector), and Y contains the labels/relevance for the training data (see below).

[W, Xi, D] = mlr_train(X, Y, C,...)

   X = d*n data matrix
   Y = either n-by-1 label of vectors
       OR
       n-by-2 cell array where 
           Y{q,1} contains relevant indices for q, and
           Y{q,2} contains irrelevant indices for q

   C >= 0 slack trade-off parameter (default=1)

   W   = the learned metric, i.e., the inner product matrix of the learned 
            space can be computed by X' * W * X
   Xi  = slack value on the learned metric (see [1])
   D   = diagnostics

By default, MLR optimizes for Area Under the ROC Curve (AUC). This can be changed by setting the "LOSS" parameter to one of several ranking loss measures:

[W, Xi, D] = mlr_train(X, Y, C, LOSS) where LOSS is one of: 'AUC': Area under ROC curve (default) 'KNN': KNN accuracy* 'Prec@k': Precision-at-k 'MAP': Mean Average Precision 'MRR': Mean Reciprocal Rank 'NDCG': Normalized Discounted Cumulative Gain

*Note:  KNN is correct only for binary classification problems; in 
        practice, Prec@k is usually a better alternative.

For KNN/Prec@k/NDCG, a threshold k may be set to determine the truncation of the ranked list. Thiscan be done by setting the k parameter:

[W, Xi, D] = mlr_train(X, Y, C, LOSS, k) where k is the number of neighbors for Prec@k or NDCG (default=3)

By default, MLR regularizes the metric W by the trace, i.e., the 1-norm of the eigenvalues. This can be changed to one of several alternatives:

[W, Xi, D] = mlr_train(X, Y, C, LOSS, k, REG) where REG defines the regularization on W, and is one of: 0: no regularization 1: 1-norm: trace(W) (default) 2: 2-norm: trace(W' W) 3: Kernel: trace(W X), assumes X is square and positive-definite

The last setting, "kernel", is appropriate for regularizing metrics learned
from kernel matrices.

W corresponds to a linear projection matrix (rotation and scaling). To learn a restricted model which may only scale, but not rotate the data, W can be constrained to diagonal matrices by setting the "Diagonal" parameter to 1:

[W, Xi, D] = mlr_train(X, Y, C, LOSS, k, REG, Diagonal) Diagonal = 0: learn a full d-by-d W (default) Diagonal = 1: learn diagonally-constrained W (d-by-1)

Note: the W returned in this case will be the d-by-1 vector corresponding
      to the main diagonal of a full metric, not the full d-by-d matrix.

Finally, we provide a stochastic gradient descent implementation to handle large-scale problems. Rather than estimating gradients from the entire training set, this variant uses a random subset of size B (see below) at each call to the cutting plane subroutine. This results in faster, but less accurate, optimization:

[W, Xi, D] = mlr_train(X, Y, C, LOSS, k, REG, Diagonal, B) where B > 0 enables stochastic optimization with batch size B

TESTING

Once a metric has been trained by "mlr_train", you can evaluate performance across all measures by using the "mlr_test" function:

Perf = mlr_test(W, test_k, Xtrain, Ytrain, Xtest, Ytest)

   W       = d-by-d positive semi-definite matrix
   test_k  = vector of k-values to use for KNN/Prec@k/NDCG
   Xtrain  = d-by-n matrix of training data
   Ytrain  = n-by-1 vector of training labels
               OR
             n-by-2 cell array where
               Y{q,1} contains relevant indices (in 1..n) for point q
               Y{q,2} contains irrelevant indices (in 1..n) for point q
   Xtest   = d-by-m matrix of testing data
   Ytest   = m-by-1 vector of training labels, or m-by-2 cell array 
               If using the cell version, indices correspond to
               the training set, and must lie in the range (1..n)

The output structure Perf contains the mean score for: AUC, KNN, Prec@k, MAP, MRR, NDCG, as well as the effective dimensionality of W, and the best-performing k-value for KNN, Prec@k, and NDCG.

For information retrieval settings, consider Xtrain/Ytrain as the corpus and Xtest/Ytest as the queries.

By testing with Wnative = [] you can quickly evaluate the performance of the native (Euclidean) metric.

MULTIPLE KERNEL LEARNING

As of version 1.1, MLR supports learning (multiple) kernel metrics, using the method described in

[3] Galleguillos, Carolina, McFee, Brian, Belongie, Serge, and Lanckriet, G.R.G.
From region similarity to category discovery. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011.

For m kernels, input data must be provided in the form of an n-by-n-by-m matrix K, where K(:,:,i) contains the i'th kernel matrix of the training data. Training the metric is similar to the single-kernel case, except that the regularization parameter should be set to 3 (kernel regularization), eg:

>> [W, Xi, D] = mlr_train(K, Y, C, 'auc', 1, 3);

returns an n-by-n-by-m matrix W which can be used directly with mlr_test. Multiple kernel learning also supports diagonally constrained learning, eg:

>> [W, Xi, D] = mlr_train(K, Y, C, 'auc', 1, 3, 1);

returns an n-by-m matrix W, where W(:,i) is the main diagonal of the i'th kernels metric. Again, this W can be fed directly into mlr_test.

For testing with multiple kernel data, Ktest should be an n-by-nTest-by-m matrix, where K(:,:,i) is the training-by-test slice of the i'th kernel matrix.

FEEDBACK

Please send any bug reports, source code contributions, etc. to

Brian McFee bmcfee@cs.ucsd.edu