ugtm (v2.0) is a package for multidimensional space analysis based on the generative topographic mapping (GTM). A complete documentation with API reference and tutorials is available online (https://ugtm.readthedocs.io). GTM is a non-linear manifold-based dimensionality reduction method introduced by Bishop et al . The ugtm package contains an implementation of GTM, and also kernel GTM (kGTM), the kernel version of the algorithm introduced by Olier et al .
GTM maps are similar to self-organizing maps  but provide a probabilistic framework that can be used to “color” the map and generate class maps or landscapes. These colored maps can then be used to build regression and classification models. GTM regression (GTR)  and GTM classification (GTC)  algorithms are implemented in ugtm. Considering that scikit-learn  is now widely used for machine learning tasks, ugtm provides scikit-learn-compatible classes for data transformation (the eGTM transformer), classification (eGTC classifier), and regression (eGTR regressor).
Other implementations of the core algorithm of GTM are available online. The netlab package  implemented in Matlab was the first implementation with published source code. GTMapTool, a software written in Free Pascal, is available as a web application on the website of the Laboratoire de Chémoinformatique in Strasbourg (http://infochim.u-strasbg.fr/mobyle-cgi/portal.py#forms::gtmaptool). ugtm provides a Python implementation of the core GTM algorithm similar to both netlab and GTMapTool, and also includes predictive modelling frameworks for classification and regression compatible with scikit-learn. The ugtm code, open to collaboration, is freely available on GitHub (https://github.com/hagax8/ugtm) under MIT license.
Implementation and architecture
ugtm v2.0 is a package implemented in pure Python. The API reference is accessible online (https://ugtm.readthedocs.io/en/latest/api.html). The main modules and their relationships are described in Figure 1. The eGTM, eGTC and eGTR classes implemented in ugtm_sklearn inherit from scikit-learn classes TransformerMixin (eGTM), ClassifierMixin (eGTC), and RegressorMixin (eGTR).
The ugtm package can be downloaded from PyPI using pip, by typing “pip install ugtm” in a terminal. In the Python console, the package can be imported by typing “import ugtm”.
(3) eGTM: GTM algorithm
The basic GTM dimensionality reduction is implemented into the scikit-learn-compatible eGTM class. The eGTM.fit() function fits a GTM model to a data matrix (a numpy array). The eGTM.transform() function uses the fitted model to generate a 2D projection for new data (also a numpy array):
from ugtm import eGTM import numpy as np
# Generate dummy train and test sets X_train = np.random.randn(100, 50) X_test = np.random.randn(50, 50)
# eGTM transformer: fit a map using X_train and project X_test eGTM().fit(X_train).transform(X_test, model=”means”)
# Model with different hyperparameters eGTM(k=10,m=5,s=1,regul=1).fit(X_train)
The four GTM hyperparameters (regul, k, m, and s) can be tuned: k is used for tuning the GTM resolution (a GTM map is discretized into a grid of [k, k] nodes), m is the number of RBF functions (defining an [m, m] grid), s is the RBF function width factor, and regul is the regularization coefficient. Implementation details can be found in the API description.
A data point can be represented in 3 ways using GTMs: responsibilities, means and modes. Responsibilities represent the probability distribution of a data point on the map; each data point is associated with a vector of k2 responsibilities (k2 = number of nodes on the GTM grid). In neural network terminology, a responsibility vector is called a feature vector and can be seen as a processed representation of a datum. These responsibilities can be used to compute the mean position of a data point on a GTM, or its mode (the node with largest responsibility).
(4) eGTC and eGTR: classification and regression using GTM
The eGTC and eGTR classes implement GTM-based classification and regression algorithms. eGTC and eGTR algorithms are based on class maps and landscapes, which are different ways of coloring a GTM. GTM class maps are constructed using discrete labels and GTM landscapes using continuous labels. New data can be projected onto these colored maps to predict labels. A GTM landscape for the S curve dataset is shown in Figure 2, and a GTM class map for the UCI handwritten digits dataset  in Figure 3, with other data projections using t-distributed stochastic neighbor embedding (t-SNE) , multidimensional scaling (MDS)  and locally linear embedding (LLE) . The visualizations were produced using ugtm, scikit-learn  and altair .
The eGTC.fit() and eGTR.fit() functions take as input two numpy arrays: a data matrix and a corresponding label vector. The eGTC.transform() and eGTR.transform() functions return predicted outcomes as numpy arrays:
from ugtm import eGTC, eGTR import numpy as np
# Generate dummy train and test sets X_train = np.random.randn(100, 50) X_test = np.random.randn(50, 50) y_train = np.random.choice([1, 2, 3], size=100)
# eGTC: predict labels for X_test y_pred = eGTC().fit(X_train,y_train). transform(X_test)
# eGTR: predict labels for X_test y_pred = eGTR().fit(X_train,y_train). transform(X_test)
(5) External resources
The package uses the following external resources:
- scikit-learn , a machine learning library that also provides data preprocessing and statistical evaluation functions.
- numpy and scipy [13, 14, 15], the main scientific packages in Python, used here for linear algebra operations and statistics.
- matplotlib , used to construct plots.
Several examples are provided in the online documentation (https://ugtm.readthedocs.io). Unit tests were conducted using the unittest Python package. The test scripts are available on GitHub (https://github.com/hagax8/ugtm/tree/master/tests). The main test scripts are the core GTM test (test_ugtm_gtm.py), the workflow test (test_ugtm_workflow.py), and scikit-learn compatibility tests (test_ugtm_sklearn.py):
- Core GTM algorithm (test_ugtm_gtm.py): the core GTM test checks matrix dimensions, convergence of the log likelihood function, and the projection of new data on the map. If the training set and test set are the same, responsibilities of the training and test sets should also be the same.
- Scikit-learn-compatible classes (test_ugtm_sklearn.py): these tests check the output dimensions of the eGTM transformer, eGTC classifier and eGTR regressor. It also checks for correct projection of new data on the GTM map.
- Workflow test (test_ugtm_workflow.py): the workflow test script was designed to test all possible workflows, for label-free data, categorical labels, and continuous labels.
Supplementary tests were implemented for plots (test_ugtm_plot.py), printing results (test_ugtm_write.py), classification models (test_ugtm_GTC.py), regression models (test_ugtm_GTR.py), and kernel algorithm (test_ugtm_kgtm.py).
These tests were carried out with Python 2.7.14 and Python 3.4.6 on Scientific Linux 6.6 and macOS High Sierra 10.13.2.
ugtm was tested on Scientific Linux 6.6 and macOS High Sierra 10.13.2 but not on Windows. ugtm is written in pure Python and should be available on any operating system supporting Python frameworks.
Python >= 2.7 (tested on Python 3.4.6 and Python 2.7.14).
Additional system requirements
ugtm does not require any supplementary data. The amount of required active memory depends on the input data and on the map size (hyperparameters k and m).
scikit-learn >= 0.20
numpy >= 1.13.1
matplotlib >= 2.0.2
scipy >= 0.19.1
mpld3 >= 0.3
List of contributors
Héléna A. Gaspar
Name: ugtm v2.0.0
Publisher: Héléna A. Gaspar
Version published: 2.0.0
Date published: 15/11/2018
Date published: 15/11/2018
ugtm was developed in English.
(3) Reuse potential
Support for ugtm is available on GitHub (https://github.com/hagax8/ugtm) – users can post issues through the GitHub platform (https://github.com/hagax8/ugtm/issues) or contribute directly to the code.
The eGTM (data transformation), eGTC (classification) and eGTR (regression) classes implemented in ugtm are fully compatible with scikit-learn and can be used in scikit-learn pipelines for visualization, regression or classification. GTM hyperparameters can be optimized using scikit-learn grid search for regression and classification tasks – examples are provided in the online documentation (https://ugtm.readthedocs.io).
Now that we have access to very large amounts of data, dimensionality reduction methods are becoming more and more popular. It is often necessary to get an overview of a dataset that exists in a space of hundreds or thousands of dimensions (data features). GTM can be a nice alternative to other dimensionality reduction algorithms such as t-SNE, MDS or LLE. It also provides a probabilistic framework that can be used to obtain a comprehensive overview of a dataset. For example, GTM can be useful to visualize and cluster multidimensional data for health research, and investigate very large datasets in chemistry or in genomics – single cell data, genotypes, endophenotypes, or polygenic risk scores. At the moment, t-SNE is very popular for these purposes but presents a major drawback: new data cannot be easily projected onto a pre-trained t-SNE map. Another possible application for GTM could be the visualization of feature vectors from deep neural networks. Hopefully, ugtm should make it easier to use GTM for these applications.
ugtm also provides opportunities for further developments. Future enhancements could include the implementation of a mini-batch version (to process data block by data block) – for now, the entire data matrix is processed in one batch. A multispace version of GTM (Stargate GTM ) could also be added in the future, as well as data projection functions for the kernel GTM algorithm. Support for other probability distributions could also be included.