Consider all tags from StackOverflow: Java, C#, multiprocessing, etc. Let's build a graph, where each vertex is a tag. Edges are represented by questions: each question connects all pairs of tags that are assigned to the question. Let's say that
w(tag_1, tag_2) = number of edges connecting tag_1 and tag_2
The more w(t1,t2) is, the better - tags are "similar" in that sense.

Given such structure, plot all tags on a 2D map using t-SNE algorithm. We want to observe relationships between tags visually.

• Obtain data about StackOverflow tags;
• create an adjacency matrix w[i][j] with weights between tags;
• feed the matrix to one of t-SNE implementations;
• plot the result nicely.

• There are around 50k tags on SO, and more than 10M questions;
• gathering the latest data from SO is not straightforward;
• t-SNE does not work with graphs out-of-the-box;
• the whole thing must run in less than 24 hours (i.e. plotting a single dump).