In [1]:
%pylab inline
t-SNE is a tool for data visualization. It reduces the dimensionality of data to 2 or 3 dimensions so that it can be plotted easily. Local similarities are preserved by this embedding.
t-SNE converts distances between data in the original space to probabilities. First, we compute conditional probabilites
$$p_{j|i} = \frac{\exp{(-d(\boldsymbol{x}_i, \boldsymbol{x}_j) / (2 \sigma_i^2)})}{\sum_{i \neq k} \exp{(-d(\boldsymbol{x}_i, \boldsymbol{x}_k) / (2 \sigma_i^2)})}, \quad p_{i|i} = 0,$$which will be used to generate joint probabilities
$$p_{ij} = \frac{p_{j|i} + p_{i|j}}{2N}.$$The $\sigma_i$ will be determined automatically. This procedure can be influenced by setting the perplexity of the algorithm.
A heavy-tailed distribution will be used to measure the similarities in the embedded space
$$q_{ij} = \frac{(1 + ||\boldsymbol{y}_i - \boldsymbol{y}_j)||^2)^{-1}}{\sum_{k \neq l} (1 + ||\boldsymbol{y}_k - \boldsymbol{y}_l)||^2)^{-1}},$$and then we minimize the Kullback-Leibler divergence
$$KL(P|Q) = \sum_{i \neq j} p_{ij} \log \frac{p_{ij}}{q_{ij}}$$between both distributions with gradient descent (and some tricks). Note that the cost function is not convex and multiple runs might yield different results.
More information can be found in these resources and in the documentation from t-SNE:
In [2]:
from sklearn.manifold import TSNE
help(TSNE)
In [3]:
from sklearn.datasets import load_iris
from sklearn.decomposition import PCA
iris = load_iris()
X_tsne = TSNE(learning_rate=100).fit_transform(iris.data)
X_pca = PCA().fit_transform(iris.data)
t-SNE can help us to decide whether classes are separable in some linear or nonlinear representation. Here we can see that the 3 classes of the Iris dataset can be separated quite easily. They can even be separated linearly which we can conclude from the low-dimensional embedding of the PCA.
In [4]:
figure(figsize=(10, 5))
subplot(121)
scatter(X_tsne[:, 0], X_tsne[:, 1], c=iris.target)
subplot(122)
scatter(X_pca[:, 0], X_pca[:, 1], c=iris.target)
Out[4]:
In high-dimensional and nonlinear domains, PCA is not applicable any more and many other manifold learning algorithms do not yield good visualizations either because they try to preserve the global data structure.
In [5]:
from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import TfidfVectorizer
categories = ['alt.atheism', 'talk.religion.misc', 'comp.graphics', 'sci.space']
newsgroups = fetch_20newsgroups(subset="train", categories=categories)
vectors = TfidfVectorizer().fit_transform(newsgroups.data)
In [ ]:
print(repr(vectors))
For high-dimensional sparse data it is helpful to first reduce the dimensions to 50 dimensions with TruncatedSVD and then perform t-SNE. This will usually improve the visualization.
In [7]:
from sklearn.decomposition import TruncatedSVD
X_reduced = TruncatedSVD(n_components=50, random_state=0).fit_transform(vectors)
In [8]:
X_embedded = TSNE(n_components=2, perplexity=40, verbose=2).fit_transform(X_reduced)
In [9]:
fig = figure(figsize=(10, 10))
ax = axes(frameon=False)
setp(ax, xticks=(), yticks=())
subplots_adjust(left=0.0, bottom=0.0, right=1.0, top=0.9,
wspace=0.0, hspace=0.0)
scatter(X_embedded[:, 0], X_embedded[:, 1],
c=newsgroups.target, marker="x")
Out[9]:
In [10]:
from sklearn.datasets import fetch_mldata
# Load MNIST dataset
mnist = fetch_mldata("MNIST original")
X, y = mnist.data / 255.0, mnist.target
# Create subset and reduce to first 50 dimensions
indices = arange(X.shape[0])
random.shuffle(indices)
n_train_samples = 5000
X_pca = PCA(n_components=50).fit_transform(X)
X_train = X_pca[indices[:n_train_samples]]
y_train = y[indices[:n_train_samples]]
In [11]:
# Plotting function
matplotlib.rc('font', **{'family' : 'sans-serif',
'weight' : 'bold',
'size' : 18})
matplotlib.rc('text', **{'usetex' : True})
def plot_mnist(X, y, X_embedded, name, min_dist=10.0):
fig = figure(figsize=(10, 10))
ax = axes(frameon=False)
title("\\textbf{MNIST dataset} -- Two-dimensional "
"embedding of 70,000 handwritten digits with %s" % name)
setp(ax, xticks=(), yticks=())
subplots_adjust(left=0.0, bottom=0.0, right=1.0, top=0.9,
wspace=0.0, hspace=0.0)
scatter(X_embedded[:, 0], X_embedded[:, 1],
c=y, marker="x")
if min_dist is not None:
from matplotlib import offsetbox
shown_images = np.array([[15., 15.]])
indices = arange(X_embedded.shape[0])
random.shuffle(indices)
for i in indices[:5000]:
dist = np.sum((X_embedded[i] - shown_images) ** 2, 1)
if np.min(dist) < min_dist:
continue
shown_images = np.r_[shown_images, [X_embedded[i]]]
imagebox = offsetbox.AnnotationBbox(
offsetbox.OffsetImage(X[i].reshape(28, 28),
cmap=cm.gray_r), X_embedded[i])
ax.add_artist(imagebox)
In [12]:
X_train_embedded = TSNE(n_components=2, perplexity=40, verbose=2).fit_transform(X_train)
In [13]:
plot_mnist(X[indices[:n_train_samples]], y_train, X_train_embedded,
"t-SNE", min_dist=20.0)
There are some modifications of t-SNE that already have been published. A huge disadvantage of t-SNE is that it scales quadratically with the number of samples ($O(N^2)$) and the optimization is quite slow. These issues and more have been adressed in the following papers: