Multipoint t-SNE

This is a Python implementation of multipoint extension described in the paper Multipoint Nebighbor Embedding, based on scikit-learn. It enables embedding individual datapoints as sets of datapoints, useful in portraying datapoints of multiple relationships, e.g., polysemic words, pictures of multiple classes, etc.

As this is experimental code, feel free to submit features and bugfixes.

Requirements

• NumPy
• scikit-learn
• SciPy
• Theano

Cython extensions need to be compiled beforehand

python setup.py build_ext

Usage

Syntax is similar to that of scikit-learn TSNE. Copies of points are introduced in stages. For this reason, the crucial thing is the schedule expressed as a list of tuples. Each tuple holds (<exaggerate>, <num-iterations>, <copy-percentage>). For instance, vanilla t-SNE schedule is [(True, 250, 0.0), (False, 750, 0.0)].

A sample run on dataset x:

mtsne = MultipointTSNE(
    n_components=2, perplexity=50.0, early_exaggeration=4.0,
    learning_rate=3.5, optimizer='adam', initial_dims=50,
    train_schedule=[
       (True,  250, 0.0),
       (False, 200, 0.0),
       (False, 200, 5.0),
       (False, 200, 5.0),
       (False, 200, 5.0),
       (False, 200, 5.0),
       (False, 200, 5.0),
       (False, 200, 5.0),
       (False, 200, 5.0),
       (False, 500, 0.0),
    ],
    verbose=15, n_iter_check=100, random_state=1234,
    method='barnes_hut', angle=0.5, cleanup_thresh=0.15)
y = mtsne.run(x=x)

Due to the dynamic nature of the algorithm, it might be resumed:

mtsne.run_schedule([(False, 200, 10.0), (False, 500, 0.0)])

Known Bugs

During phases with many iterations, unnecessary copies get pushed infinitely far from the main cloud of points causing numerical errors.

Citation

@Inbook{Lancucki2017,
  author="Lancucki, Adrian and Chorowski, Jan",
  editor="Ek{\v{s}}tein, Kamil and Matou{\v{s}}ek, V{\'a}clav",
  title="Multipoint Neighbor Embedding",
  bookTitle="Text, Speech, and Dialogue: 20th International Conference, TSD 2017, Prague, Czech Republic, August 27-31, 2017, Proceedings",
  year="2017",
  publisher="Springer International Publishing",
  address="Cham",
  pages="456--464",
  isbn="978-3-319-64206-2",
  doi="10.1007/978-3-319-64206-2_51",
  url="https://doi.org/10.1007/978-3-319-64206-2_51"
}