https://github.com/rmitsch/tapas

Assisted hyperparameter optimization for t-SNE in a word embedding context.
https://github.com/rmitsch/tapas

dimensionality-reduction hyperparameter-optimization t-sne visualization word-embeddings

Last synced: 2 days ago
JSON representation

Assisted hyperparameter optimization for t-SNE in a word embedding context.

• Host: GitHub
• URL: https://github.com/rmitsch/tapas
• Owner: rmitsch
• License: gpl-3.0
• Created: 2017-10-27T12:59:42.000Z (about 7 years ago)
• Default Branch: master
• Last Pushed: 2018-07-13T12:16:11.000Z (over 6 years ago)
• Last Synced: 2024-12-29T10:44:31.535Z (5 days ago)
• Topics: dimensionality-reduction, hyperparameter-optimization, t-sne, visualization, word-embeddings
• Language: Python
• Homepage:
• Size: 4.01 MB
• Stars: 1
• Watchers: 4
• Forks: 0
• Open Issues: 2
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# tAPAS

### Assisted Parameter optimization by Approximating neighbourhood Similarity

Bayesian optimization of hyperparameters for t-SNE in the context of word embeddings - i. e.: Input is a word embedding (labels + coordinates in high-dimensional space). The optimization procedure samples the parameter space to generate low-dimensional approximations of the original word embeddig data using t-SNE. The quality/truthfulness of the resulting models are evaluated with several metrics:

* Trustworthiness: Measure for proportion of points too close together1 in the low-dimensional space.

* Continuity: Measure for proportion of points too far apart1 in the low-dimensional space.

* Generalization: Generalization error of 1-nearest neighbour classifier (e. g. word embedding is clustered in high-dimensional and low-dimensional space - the higher the similarity between the cluster labels, the lower the generalization error).

* Relative word embedding quality: QVEC [1] is used to evalute the intrinsic quality of the original word embedding and its dimensionality-reduced projection. The ratio is referred to as 'relative word embedding quality'.

The first three measures were chosen following [2].

![Main View](https://raw.githubusercontent.com/rmitsch/tapas/master/doc/main.png)

![Generation of New Runs](https://raw.githubusercontent.com/rmitsch/tapas/master/doc/run_generation.png)

1: In terms of neighbourhood ranks, not absolute distances.

_____

[1] Y. Tsvetkov, M. Faruqui, W. Ling, G. Lample, and C. Dyer, “Evaluation of Word Vector Representations by Subspace Alignment,” 2015, pp. 2049–2054.

[2] L. J. P. van der Maaten, E. O. Postma, and H. J. van den Herik, Dimensionality Reduction: A Comparative Review. 2008.