2
$\begingroup$

I have blood test results with 20 features from 170 patients, and trying to predict a categorical disease outcome. The input features are all continuous values except sex (0:male, 1:female) There are no missing values (used imputation for the few missing values)

When I plotted a 2D t-SNE plot, the results is as the attached image.(Figure 1) A single curved line shape. I expected the plot to be more scattered, but even if I increase the perplexity value from 13 to 20 or 30 it still remains as a line shape. (the initial value of 13 is from a recommendation that sqrt # of samples is a good starting point)

Reducing the perplexity value to 6 makes the datapoints form small chunks (Figure 2), and I am aware this is because small perplexity values make the datapoints conserve proximity with only a few neighbors so this results in small groups.

In a related question this curved line shape appeared in a time-series dataset. I undetstood that time-series data have datapoints that change in small increments across most of the features, so the t-SNE results form a line.

So this means in my dataset, patients that have similar values, the "direction" of similarity is not random due to human physiology. For instance if patient A & B are similar and B has a slightly higher RBC value, the WBC, hematocrit are also higher and BIL values are lower. Hence, each individual patient may look like data from a time-series dataset.

Does this explanation make sense? Or is could there be a different explanation of this line-shaped distribution of a t-SNE plot?

perplexity value = 12 perplexity value = 6

Improve this question
$\endgroup$

2 Answers 2

Reset to default
2
$\begingroup$

From my own experience, if you have curved lines in your t-SNE plots, it usually implies your original data are scattered similarly in lines as well. One thing you can confirm whether this hypothesis is correct is that you can do principle component analysis (PCA) on your original data and then plot the first two dimensions. Usually, this plot will also be in curved lines.

Your explanation is plausible. However, I suggest you plot your first two PCA components to see what is really going on in your data.

Improve this answer
$\endgroup$
1
$\begingroup$

There's also a helpful discussion of this topic, in the context of UMAP, over at https://github.com/lmcinnes/umap/issues/233.

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ Since links can change, summarizing what the link is about would help improve this answer :) $\endgroup$
    – m13op22
    Commented Feb 19 at 16:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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