t-viSNE: Interactive Assessment and Interpretation of t-SNE Projections
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Angelos Chatzimparmpas 1acb88ac6c responsiveness 3 months ago
css responsiveness 3 months ago
data Added two more data sets used in the paper 5 months ago
js responsiveness 3 months ago
modules Bump open from 0.0.5 to 7.0.3 in /modules/papa 5 months ago
textures fixed css 1 year ago
.gitignore Ignored cache dirs 5 months ago
LICENSE Add README.md 1 year ago
LICENSE.txt update 10 months ago
Makefile.win update 10 months ago
README.md responsiveness 3 months ago
betas.txt fix responsiveness at first panel 3 months ago
bh_tsne test 10 months ago
bhtsne.py test 10 months ago
cost_per_iter.txt fix responsiveness at first panel 3 months ago
cost_per_point.txt fix responsiveness at first panel 3 months ago
fast_tsne.m update 10 months ago
index.html responsiveness 3 months ago
ps new 5 months ago
requirements.txt Fixed requirements 5 months ago
result.dat fix responsiveness at first panel 3 months ago
run.sh new 10 months ago
sptree.cpp update 10 months ago
sptree.h update 10 months ago
thumbnail_representative.png new pic 5 months ago
tsne.cpp test 10 months ago
tsne.h test 10 months ago
tsneGrid.py new 5 months ago
tsne_main.cpp test 10 months ago
vptree.h update 10 months ago

README.md

t-viSNE: Interactive Assessment and Interpretation of t-SNE Projections

Codacy Badge

This Git repository contains the code that accompanies the research paper “t-viSNE: Interactive Assessment and Interpretation of t-SNE Projections”. The details of the experiments and the research outcome are described in the paper.

Note: t-viSNE is optimized to work better for standard resolutions (such as 1440p/QHD (Quad High Definition) and 1080p). Any other resolution might need manual adjustment of your browser’s zoom level to work properly.

Note: The tag paper-version matches the implementation at the time of the paper’s publication. The current version might look significantly different depending on how much time has passed since then.

Note: This software is based on the bhtsne library, its native executable and the python interface that is used to call the native executable. This library is the official implementation of t-SNE, made by its authors. Using the exact same input data, different systems will generate slightly different outputs in this library, and such differences will propagate to our software.

Data Sets

All data sets used in the paper are in the data folder, formatted as comma separated values (csv). Most of them are available online from the UCI Machine Learning Repository: Iris, Breast Cancer Wisconsin (Original), Pima Indians Diabetes, and SPECTF. We also used a custom-made data set with Gaussian clusters.

Requirements

For the backend:

You can install all the backend requirements with the following command:

pip install -r requirements.txt

For the frontend:

There is no need to install anything for the frontend, since all modules are in the repository.

Usage

Below is an example of how you can get t-viSNE running using Python for both frontend and backend. The frontend is written in Javascript/HTML, so it could be hosted in any other web server of your preference. The only hard requirement (currently) is that both frontend and backend must be running on the same machine.

# first terminal: hosting the visualization side (client)
# for Python3
python3 -m http.server 

or

# for Python2
python -m SimpleHTTPServer 8000
# second terminal: hosting the computational side (server)
FLASK_APP=tsneGrid.py flask run

Then, open your browser and point it to localhost:8000. We recommend using an up-to-date version of Google Chrome.

Reproducibility of the Results

The following instructions describe how to reach the results present in Figure 1 of the article. The aforementioned figure is connected with the Subsection 5.2 (Use Case: Improving Diabetes Classification) and is the main use case described in the paper.

Note: We used OSX and Google Chrome in all our tests, so we cannot guarantee that it works in other OS or browser. However, since t-viSNE is written in JS and Python, it should work in all the most common platforms.

Tip: The Reset Filters button illustrated in Figure 1(h), resets all the applied interactions in case you made a mistake and you want to redraw something.

  • Step 1: Make sure the “Pima Indian Diabetes” data set is selected (top-left corner), then press the Execute new t-SNE analysis button that is shown in Figure 1(a).
  • Step 2: From the new pop-up tab named as t-SNE Grid Search Resulting Diverse Projections, you have to select the Continuity © quality metric. Please note that the first time you execute the analysis and, consequently, run the hyper-parameter search, it will take a few minutes before the different options of projections show up in the grid. After the first time, the results are cached and will be re-used to make the process faster.
  • Step 3: From the five projections seen on the top row (out of the 25), choose the fourth projection with Quality Metrics Average (QMA) approximately 0.43 and rather high Continuity © (around 0.69) by clicking on one of its points (you will know it is selected when the border becomes red). Press Confirm and wait for the system to load all the views. The t-SNE overview is the same as in Figure 7(a) and the main projection can be seen in Figure 7(a). Also, Figures 7(b) and 7(d) show the same representations as you should have in your screen now.
  • Step 4: To reproduce Figure 7(e), you have to use the Group Selection option from the Interaction Modes illustrated in Figure 1(h). Carefully select the areas depicted in red color in Figure 7© by holding the left mouse button and drawing a shape with the lasso interaction provided by the tool. Note that the lasso only works correctly if you release the mouse near the starting point. The first area is “High remaining cost”, the second is “Low-density ”tip””, and the third “Low-density cluster”. Please note that the results in Figure 7(e) start from 20 to 40 k-values of nearest neighbors.
  • Step 5: For Figure 7(g), you use the Dimension Correlation option from the Interaction Modes illustrated in Figure 1(h). You can draw the line seen in Figure 7(g) and Figure 7© in blue by left-clicking to position each point in sequence, one after the other. When it’s done, with the right mouse click you confirm the drawn shape and observe in the Dimension Correlation view depicted in Figure 1(j) the values of Figure 7(f). Here, do not forget to adjust the correlation threshold in order to cover the entire cluster. Finally, if you click the bar for the “Insulin” dimension you get Figure 7(g). The exact values of “Insulin” are now shown in the legend of Figure 1(i) instead of the Density and the Remaining Cost as it was before in Figure 7(c, bottom-right corner).
  • Step 6: To receive the image shown in Figure 1, you choose with the Group Selection described in Step 4 the third cluster named as “Low-density cluster” (cf. Figure 7©). Then you should click the Optimize Selection button (see Figure 1(e)), which is going to resort all the projections.
  • Step 7: The next step is to choose the Quality Metrics Average (QMA) option instead of Continuity © with the dropdown selection seen in Figure 1(e). This will sort the projections based on the average value of all the provided quality metrics. Afterward, you click and choose the first projection out of the six shown in Figure 1(e). Now Figure 1 is loaded for you.
  • Step 8: Finally, you use the Dimension Correlation option from the Interaction Modes illustrated in Figure 1(h) and draw a line (i.e., click to place two individual points) as shown in Figure 1(f). Next, press right click to confirm the drawn shape as in Step 5. If the line is drawn correctly, then you just go to the Visual Mapping panel (cf. Figure 1(i)) and you select KNN instead of Distance for the Correlation measurement. After this modification, you will find the K-value (KNN) field which should be set to 34 (see Figure 1(i)).

Outcome: The above process describes how you will be able to reproduce precisely the results presented in Figures 1 and 7 of the paper. Thank you for your time!

Corresponding Author

For any questions with regard to the implementation or the paper, feel free to contact Angelos Chatzimparmpas.