2
$\begingroup$

I have about a half-dozen variables, each of which can have anywhere from three to ten outcomes. I have to measure the degree of separation/similarity between rows.

Either we can do some sort of weighted system for each variable and then go by a cutoff percent, or we can apply some sort of classification technique?

I was thinking k-NN, but it's not designed for categorical data. I know that there is an extension for categorical data, but what's the difference between percentage/weighted similarity?

Open to thoughts or clarifications of my assumptions. Thank you.

EDIT: Adding some example data (not exactly the same but simplified)

TIME RANGE |   TYPE   |  REGION  |   MANAGER   | ....... | ....... |
===========|==========|==========|=============|=========|=========|
  SHORT    |   STOCK  |    US    |  JP MORGAN  |   ...   |   ...   |
  SHORT    |   STOCK  |   INTL   |     BOA     |   ...   |   ...   |
  LONG     |   BOND   |    US    |   GOLDMAN   |   ...   |   ...   |
  SHORT    |   BOND   |    US    |     BOA     |   ...   |   ...   |
  MEDIUM   |   BOND   |   INTL   |     BOA     |   ...   |   ...   |
  LONG     |   STOCK  |   INTL   |  JP MORGAN  |   ...   |   ...   |
  MEDIUM   |   BOND   |   INTL   |   GOLDMAN   |   ...   |   ...   |
  SHORT    |   BOND   |    US    |   GOLDMAN   |   ...   |   ...   |
  MEDIUM   |   BOND   |    US    |   GOLDMAN   |   ...   |   ...   |
  MEDIUM   |   STOCK  |   INTL   |   GOLDMAN   |   ...   |   ...   |

EDIT: Domain knowledge, want to find out how similar each row was to another. So we would have some sort of key/identifier, and we would only want to match two of them. I guess we can assign some sort of coded rating to each row? But if we are just comparing two specific rows, then how would we go about it?

EDIT: I believe this would be some sort of unsupervised learning. Hence, the dependent variable would not necessarily be categorical. The independent ones WOULD be.

$\endgroup$
  • $\begingroup$ Can you show us some data and explain little bit more by what you mean with ``degree of separation/similarity between rows''? $\endgroup$ – Vladislavs Dovgalecs Apr 1 '15 at 16:32
  • $\begingroup$ Hi. Added it. Please let me know if it helps. Thanks. $\endgroup$ – vdiddy Apr 1 '15 at 16:47
  • $\begingroup$ Thanks for the data. Would you like to classify the rows in some predefined classes (assumes you have labels) or cluster the data? $\endgroup$ – Vladislavs Dovgalecs Apr 1 '15 at 16:51
  • 1
    $\begingroup$ Can you also provide some domain knowledge, e.g. how would you read this data and classify/cluster it? How would you group a handful of rows? $\endgroup$ – Vladislavs Dovgalecs Apr 1 '15 at 17:01
  • $\begingroup$ So basically we want to see how "similar" each of the rows are. I mean we could just do percentages (3/5 match so it's 60%), etc. but it doesn't seem to add much technical weight. $\endgroup$ – vdiddy Apr 1 '15 at 17:18
1
$\begingroup$

Multinomial logit is a technique to classify purely categorical dependent variables, and permits some measure of distance between outcomes. But I doubt this is an exhaustive list, simply one I am somewhat familiar with.

$\endgroup$

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.