4
$\begingroup$

I apologize for lack of terminology, I'm no computer scientist. I have a problem of validating paths in a directed graph with complex nodes.

The full description is the following:

  1. I have a decent set (about 1K) of directed graphs;
  2. Each node contains a complex data structure (it is a hierarchical data structure, not a picture or sound);
  3. I have some of paths in those graphs known as "correct" paths (based mostly on data in nodes);
  4. And I have some of paths in those graphs known as "incorrect" paths (with classification why it is incorrect).

I'd like to predict given a graph with those complex nodes and a path, is this path "correct".

Which machine learning algorithm will suit me best? In general, what approach I should use?

Edit:

  1. Each full graph is either have app paths processed (correct/incorrect) or completely blank (no path is processed);
  2. Correctness depends on both position of node in a graph AND data in the node;
  3. Humans would need heuristics to decide or guess which paths are correct;
  4. Most of the paths are "correct";
  5. I hope to convert human heuristics to some kind of "correctness" recognition.
$\endgroup$
5
  • $\begingroup$ "correctness" depends on only on data in nodes but also on position of the node in the graph. There is a "correctness" flag assigned externally humans could figure it out based on the data but with a VERY large amount of effort. Again, there is an external process which tells if data is correct for the whole path or if it is incorrect. Data about correctness is available for the whole graph (all paths) or not available at all. I don't know if it can be used to predict correctly, I'd like to play and find out. :) $\endgroup$
    – Artem
    Feb 7 '16 at 17:48
  • $\begingroup$ Based on correctness of some paths in graph you might not be able to predict correctness of others. it is more about the whole graph $\endgroup$
    – Artem
    Feb 7 '16 at 17:51
  • $\begingroup$ It is extremely hard, it would be like automating image recognition. It is not just logical, it is good amount of heuristics. When I said possible, humans can work with graphical representation of this data, but to figure out from those data structures what is going on might take months and still need lots of human heuristics on how it should work. I don't know, it might be related to search, in this case I'm completely lost and don't understand what do I need to do here. $\endgroup$
    – Artem
    Feb 7 '16 at 18:28
  • $\begingroup$ I hope to convert human heuristics to some kind of "correctness" recognition. Does it sound like neural networks or something like that? $\endgroup$
    – Artem
    Feb 7 '16 at 18:29
  • 1
    $\begingroup$ If the humans are using something more like intuition, then yes an ML technique might help. Often it's a good indicator to use ML when a human expert can do something, but it is hard to identify the logic. I suggest putting some of these salient points from your comments into the question, to help someone answering. $\endgroup$ Feb 7 '16 at 18:48
3
$\begingroup$

This is a good question but it's rather complicated. I can suggest two approaches:

$\endgroup$
2
$\begingroup$

A third approach: Graph analytics, is the approach which lets you measure how strong relationships between subjects are, when represented as graphs.

Graph Analytics

$\endgroup$

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.