2
$\begingroup$

Given a data set of N features, wherein some the features in this set were derived from other features from the same set, I am trying to discover inter dependencies between features (something like this Input feature(s) -> output feature(s)).

Note that,there can be multiple dependencies in the same feature set. Can someone suggest some technique to approach this problem.

$\endgroup$

3 Answers 3

0
$\begingroup$

For me it seems you are looking for Association Rule Learning. A quite simple example for an algorithm you could use is the Apriori Algorithm. You could try that one first and see if it does what you want to achieve. Then you can take a look into more sophisticated algorithms.

Generally these algorithms try to find associations between your features. E.g. they are able to find the implication, that many people who buy Coca Cola & Beer, will buy a pack of chips - if supported by the data:

$(Coca Cola, Beer)\implies(Chips)$.

$\endgroup$
3
  • $\begingroup$ Thanks for replying. AR could be an option but we ruled it out for now, as it works at value level and thus may create too many or two few rules (even when the 2 features have some interdependence). $\endgroup$
    – user116948
    Commented Sep 12, 2018 at 11:47
  • $\begingroup$ what is your data and how are the corresponding features composed? $\endgroup$
    – André
    Commented Sep 12, 2018 at 12:17
  • $\begingroup$ The data is from banking domain, which has categorical and numeric features. This data was created through a set of complex rules, and not all features were used in each rule (that means there is not one single Target/dependent feature). For example a feature 'Final Disbursement' is derived from 4 other numeric and categorical features, some of those were in turn derived from some other features from the set. All in all, it's a jumbled data set and we want to discover what might have created what (or the clear dependencies). $\endgroup$
    – user116948
    Commented Sep 12, 2018 at 12:58
0
$\begingroup$

Tree-based methods (e.g. random forests) and boosted tree methods (e.g. XGBoost) are usually quite good at detecting underlying relationships between features, and it's usually quite straightforward to extract measures of feature "importance".

$\endgroup$
1
  • $\begingroup$ Well, your answer is informative, and could be more question specific by being general about exploring inter dependencies among features. $\endgroup$
    – naive
    Commented Sep 12, 2018 at 17:42
0
$\begingroup$

You could try Principal Components Analysis, or a somewhat more informative alternative: least-squares feature selection.

$\endgroup$

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