Metric learning algorithms in Python
metric-learn is an open source Python package implementing supervised and weakly-supervised distance metric learning algorithms. As part of scikit-learn-contrib, it provides a unified interface compatible with scikit-learn which allows to easily perform cross-validation, model selection, and pipelining with other machine learning estimators. metric-learn is thoroughly tested and available on PyPi under the MIT licence.READ FULL TEXT VIEW PDF
This paper describes the discipline of distance metric learning, a branc...
Deep metric learning algorithms have a wide variety of applications, but...
In recent years, we have witnessed a surge of interests in learning a
Recent work in metric learning has significantly improved the
In machine learning, the choice of a learning algorithm that is suitable...
Traditionally, the automatic recognition of human activities is performe...
stream-learn is a Python package compatible with scikit-learn and develo...
Metric learning algorithms in Python
Many approaches in machine learning require a measure of distance between data points. Traditionally, practitioners would choose a standard distance metric (Euclidean, City-Block, Cosine, etc.) using a priori knowledge of the domain. However, it is often difficult to design metrics that are well-suited to the particular data and task of interest. Distance metric learning, or simply metric learning (Bellet15)
, aims at automatically constructing task-specific distance metrics from data. A key advantage of metric learning is that it can be applied beyond the standard supervised learning setting (data points associated with labels), in situations where only weaker forms of supervision are available (e.g., pairs of points that should be similar/dissimilar). The learned distance metric can be used to perform retrieval tasks such as finding elements (images, documents) of a database that are semantically closest to a query element. It can also be plugged into other machine learning algorithms, for instance to improve the accuracy of nearest neighbors models (for classification, regression, anomaly detection…) or to bias the clusters found by clustering algorithms towards the intended semantics. Finally, metric learning can be used to perform dimensionality reduction. These use-cases highlight the importance of integrating metric learning with the rest of the machine learning pipeline and tools.
metric-learn is an open source package for metric learning in Python, which implements many popular metric-learning algorithms with different levels of supervision through a unified interface. Its API is compatible with scikit-learn (scikit-learn), a prominent machine learning library in Python. This allows for streamlined model selection, evaluation, and pipelining with other estimators.
Metric learning is generally formulated as an optimization problem where one seeks to find the parameters of a distance function that minimize some objective function over the input data. All algorithms currently implemented in metric-learn learn so-called Mahalanobis distances. Given a real-valued parameter matrix of shape (n_components, n_features) where n_features is the number of features describing the data, the associated Mahalanobis distance between two points and is defined as
. This is equivalent to Euclidean distance after linear transformation of the feature space defined by. Thus, if
is the identity matrix, standard Euclidean distance is recovered. Mahalanobis distance metric learning can thus be seen as learning a new embedding space, with potentially reduced dimensionn_components. Note that can also be written as , where we refer to as the Mahalanobis matrix.
Metric learning algorithms can be categorized according to the form of data supervision they require to learn a metric. metric-learn currently implements algorithms that fall into the following categories. Supervised learners learn from a dataset with one label per training example, aiming to bring together points from the same class while spreading points from different classes. For instance, data points could be face images and the class could be the identity of the person (see Figure 0(a)). Pair learners require a set of pairs of points, with each pair labeled to indicate whether the two points are similar or not. These methods aim to learn a metric that brings pairs of similar points closer together and pushes pairs of dissimilar points further away from each other. Such supervision is often simpler to collect than class labels in applications when there are many labels. For instance, a human annotator can often quickly decide whether two face images correspond to the same person (Figure 0(b)) while matching a face to its identity among many possible people may be difficult. Finally, quadruplet learners consider 4-tuples of points and aim to learn a metric that brings the two first points of each quadruplet closer than the two last points. This can be used to learn a metric space in which closer points are more similar with respect to an attribute of interest, which may be continuous and difficult to annotate accurately (e.g., the age of a person on an image, see Figure 0(c)). Quadruplet supervision is also used in problems with a class hierarchy.
The current release of metric-learn (v.0.5.0) can be installed from the Python Package Index (PyPI), for Python 2.7 and 3.5 or later. The source code is available on GitHub at http://github.com/scikit-learn-contrib/metric-learn and is free to use, provided under the MIT license. metric-learn depends on core libraries from the SciPy ecosystem: numpy, scipy, and scikit-learn. Detailed documentation (including installation guidelines, the description of the algorithms and the API, as well as examples) is available at http://contrib.scikit-learn.org/metric-learn. The development is collaborative and open to all contributors through the usual GitHub workflow of issues and pull requests. Community interest for the package has been demonstrated by its recent inclusion in the scikit-learn-contrib organization which hosts high-quality scikit-learn-compatible projects,111https://github.com/scikit-learn-contrib/scikit-learn-contrib and by its more than 740 stars and 170 forks on GitHub at the time of writing. The quality of the code is ensured by a thorough test coverage (96% as of July 2019). Every new contribution is automatically checked by a continuous integration platform to enforce sufficient test coverage as well as syntax formatting with flake8.
Currently, metric-learn implements 9 popular metric learning algorithms. Supervised learners include Neighborhood Components Analysis (NCA, Goldberger04), Large Margin Nearest Neighbors (LMNN, Weinberger09), Relative Components Analysis (RCA, Shental02),222RCA takes as input slightly weaker supervision in the form of chunklets (groups of points of same class). Local Fisher Discriminant Analysis (LFDA, Sugiyama07) and Metric Learning for Kernel Regression (MLKR, Weinberger07). The latter is designed for regression problems with continuous labels. Pair learners include Mahalanobis Metric for Clustering (MMC, Xing2002a), Information Theoretic Metric Learning (ITML, Davis07) and Sparse High-Dimensional Metric Learning (SDML, Qi09). The package implements one quadruplet learner: Metric Learning from Relative Comparisons by Minimizing Squared Residual (LSML, Liu12). Detailed descriptions of these algorithms can be found in the package documentation.
metric-learn provides a unified interface to all metric learning algorithms. It is designed to be fully compatible with the functionality of scikit-learn. All metric learners inherit from an abstract BaseMetricLearner class, which itself inherits from scikit-learn’s BaseEstimator. All classes inheriting from BaseMetricLearner should implement two methods: get_metric (returning a function that computes the distance, which can be plugged into scikit-learn estimators like KMeansClustering) and score_pairs (returning the distances between a set of pairs of points passed as a 3D array). Mahalanobis distance learning algorithms also inherit from a MahalanobisMixin interface, which has an attribute components_ corresponding to the transformation matrix of the Mahalanobis distance. MahalanobisMixin implements get_metric and score_pairs accordingly as well as a few additional methods. In particular, transform allows to transform data using components_, and get_mahalanobis_matrix returns the Mahalanobis matrix .
Supervised metric learners inherit from scikit-learn’s base class TransformerMixin, the same base class used by sklearn.LinearDiscriminantAnalysis and others. As such, they are compatible for pipelining with other estimators via sklearn.pipeline.Pipeline.
Weakly supervised algorithms (pair and quadruplet learners) fit and predict on a set of tuples passed as a 3-dimensional array. Tuples can be pairs or quadruplets depending on the algorithm. Pair learners take as input an array-like pairs of shape (n_pairs, 2, n_features), as well as an array-like y_pairs of shape (n_pairs,) giving labels (similar or dissimilar) for each pair. In order to predict the labels of new pairs, one needs to set a threshold on the distance value. This threshold can be set manually or automatically calibrated (at fit time or afterwards on a validation set) to optimize a given score such as accuracy or F1-score using the method calibrate_threshold. Quadruplet learners work on array-like of shape (n_quadruplets, 4, n_features), where for each quadruplet the two first elements are the ones we want to be closer than the two last ones. They can naturally predict whether a new quadruplet is in the right order by comparing the two pairwise distances. These design choices enable use of scikit-learn’s scoring functions out of the box, as well as the standard routines for model selection, including GridSearchCV. To illustrate, the following code snippet computes cross validation scores for ITML (with default parameters) on pairs from Labeled Faces in the Wild (Huang12). [fontsize=]python ¿¿¿ from sklearn.datasets import fetch_lfw_pairs ¿¿¿ from sklearn.model_selection import cross_validate, train_test_split ¿¿¿ from sklearn.decomposition import PCA ¿¿¿ from metric_learn import ITML ¿¿¿ pairs, y_pairs = [fetch_lfw_pairs()[key] for key in [’pairs’, ’target’]] ¿¿¿ pairs = PCA(n_components=25).fit_transform(pairs.reshape(4400, -1)).reshape(-1, 2, 25) ¿¿¿ pairs, _, y_pairs, _= train_test_split(pairs, 2*y_pairs-1) ¿¿¿ cross_validate(ITML(), pairs, y_pairs, scoring=’roc_auc’, return_train_score=True)
metric-learn is under active development. We list here some promising directions to further improve the package. To scale to large datasets, we would like to implement stochastic solvers (SGD and its variants), forming batches of tuples on the fly to avoid loading all data in memory at once. We also plan to incorporate new algorithms that provide added value to the package, in particular some that learn from triplet supervision (Schultz2003a), can deal with multi-label (liu15) and high-dimensional problems (Liu19), or learn other forms of metrics (e.g., nonlinear ones, bilinear similarities, and multiple local metrics, see Bellet15).
We are thankful to Inria for funding 2 years of development. We also thank scikit-learn developers from the Inria Parietal team (in particular Gaël Varoquaux, Alexandre Gramfort and Olivier Grisel) for fruitful discussions on the design of the API and funding to attend SciPy 2019, as well as scikit-learn-contrib reviewers for their valuable feedback.