Quoting from one of the authors:

t-Distributed Stochastic Neighbor Embedding (t-SNE) is a (prize-winning) technique for dimensionality reduction that is particularly well suited for the visualization of high-dimensional datasets.

So it sounds pretty great, but that is the Author talking.

Another quote from the author (re: the aforementioned competition):

What have you taken away from this competition?
Always visualize your data first, before you start to train predictors on the data! Oftentimes, visualizations such as the ones I made provide insight into the data distribution that may help you in determining what types of prediction models to try.

Information must1 be being lost -- it is a dimensionality reduction technique afterall. However, as it is a good technique to use when visualising, the information lost is less valuable than the information highlighted (/made visible/comprehend-able through reduction to 2 or 3 dimensions).

So my question is:

  • When is tSNE the wrong tool for the job?
  • What kind of datasets cause it to not function,
  • What kind of questions does it look like it can answer, but it actually can not?
  • In the second quote above it is recommended to always visualise your dataset, should this visualisation always be done with tSNE?

I expect that this question might be best answered in the converse, ie answering: When is tSNE the right tool for the job?

I have been cautioned not to rely on tSNE to tell me how easy data will be classifiable (separated into classes -- a discriminative model) The example of it being misleading was, that, for the two images below, a generative model2 was worse for the data visualised in the first/left (accuracy 53.6%) than an equivalent one for the second/right (accuracy 67.2%).

first second

1 I could be wrong about this I may sit down and try at a proof/counter example later

2 note that a generative model is not the same as a discriminitive model, but this is the example I was given.

  • 1
    $\begingroup$ Regarding your statement that "information must be being lost": A bijective mapping between sets $A$ and $B$ exists if and only if their cardinality agrees, $|A| = |B|$. And we have, for example, $|\mathbb{N}| = |\mathbb{N}^n| = \aleph_0$ and $|\mathbb{R}| = |\mathbb{R}| = |\mathbb{R}^n| = \aleph_1$ (see here). That is, in principle we can cram as much information in $\mathbb{R}^2$ as in $\mathbb{R}$. $\endgroup$ – Lucas Jan 8 '15 at 9:16
  • $\begingroup$ @Lucas: Ah, of-course. (how did I not realise that) $\endgroup$ – Lyndon White Jan 8 '15 at 9:22
  • $\begingroup$ What generative model you are trying? $\endgroup$ – WeiChing 林煒清 Jul 29 '15 at 5:38
  • $\begingroup$ @Wei-ChingLin I'm not sure what kind of generative model was being used. Likely some kind of Deep Belief Network, Deep Boltzmann Manchine, or Autoencoder. Not really relevant to the heart of the question $\endgroup$ – Lyndon White Jul 29 '15 at 6:40
  • 2
    $\begingroup$ Relevant: distill.pub/2016/misread-tsne $\endgroup$ – Lyndon White Dec 5 '16 at 4:35

T-Sne is a reduction technique that maintains the small scale structure (i.e. what is particularly close to what) of the space, which makes it very good at visualizing data separability. This means that T-Sne is particularly useful for early visualization geared at understanding the degree of data separability. Other techniques (PCA for example) leave data in lower dimensional representations projected on top of each other as dimensions disappear, which makes it very difficult to make any clear statement about separability in the higher dimensional space.

So for example, if you get a T-Sne graph with lots of overlapping data, odds are high that your classifier will perform badly, no matter what you do. Conversely, if you see clearly separated data in the T-Sne graph, then the underlying, high-dimensional data contains sufficient variability to build a good classifier.

| cite | improve this answer | |
  • 3
    $\begingroup$ That is a very good explanation of what T-SNE is, thanks. But I'm not seeing answers to my actual questions (See the the dot points in the opening post.) $\endgroup$ – Lyndon White Oct 31 '15 at 14:07
  • 5
    $\begingroup$ This does not answer the question at all. $\endgroup$ – amoeba Sep 23 '16 at 22:06

Out of the box, tSNE has a few hyperparameters, the main one being perplexity. Remember that heuristically, perplexity defines a notion of similarity for tSNE and a universal perplexity is used for all data-points. You could try generating a labelled dataset where each cluster has wildly different perplexity. This can be accomplished by doing a mixture of gaussians, with a wide range of different variances. I'm guessing this will also cause issues in the Barnes-Hut implementation of tSNE, which relies on quartiling data and using only nearest neighbours. tSNE also has an initial relaxation period, which attempts to pass clusters through each other. During this period, there is no penalty or repulsion. So for example, if your data looks a matted clump of noodles (each noodle representing a given cluster), you're gonna have a hard time calibrating the initial pass through, and I doubt tSNE will work well. In some sense I think this hints that tSNE will not work well if your data is woven together and initially resides in a low dimensional space, say 5.

In general tSNE is good because of the "t" part, which resolves an outstanding issue in SNE of how to space points in lower dimensional spaces, compared to higher dimensions. It turns out that on average, the spacing of data points in higher dimensions behaves completely differently from lower dimensions. In particular, tSNE strongly advocates against using Gaussians to measure distances in lower dimensions, opting instead for the one dimensional $t$ distribution (i.e. the Cauchy Distribution) which has heaver tails and allows for more spread in the lower dimensional representation. So conceivably the "t" in tSNE could also be a hyperparameter, where instead you can choose different distributions (albeit at a high computational cost).

You should think of tSNE as an unsupervised method of clustering, and as such there's zero reason to think that it's the only tool for the job. I think overall it can be a fantastic tool if calibrated right. However it is quite slow on large datasets and you might be better off using some optimized form of $k$-means for example, or even PCA, depending on how sparse the data is.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.