1
$\begingroup$

I have a dataset of 815 positive examples and 9492 negative examples for a certain class. Each example is represented by 12 features and a target label (i.e. TRUE/FALSE). The dataset is in a CSV file and can be downloaded from here.

The 12 features are measures of different quality aspects. The question I am trying to answer is whether or not the positive instances of this class are significantly better (or worse) in any of these quality aspects compared with the negative instances. In other words, I need to know which of the features (if any) have significantly different values in the positive/negative examples.

I would deeply appreciate it if you could suggest one or more appropriate statistical analysis for this dataset.

UPDATE: a new version of the dataset with 15 features has been uploaded.

UPDATE 2: I have tried using Mann-Whitney test and I have surprisingly found statistically significant differences in 12 out of 15 features, 11 of which are significant at p-value <0.0001. This may indicate that the test used is not the appropriate one. If it is indeed an appropriate test, how can I appropriately select the 3 features with the most significant differences?

UPDATE 3: all features are on the Ratio scale.

$\endgroup$
5
  • $\begingroup$ please be more specific as to the goal of your analysis $\endgroup$ Aug 26, 2015 at 9:41
  • $\begingroup$ Thanks for your response. The 15 features are measures of different quality aspects. The question I am trying to answer is whether or not the positive instances of this class are significantly better (or worse) in any of these quality aspects compared with the negative instances. $\endgroup$ Aug 26, 2015 at 9:52
  • $\begingroup$ The data on features in the .csv file don't all seem to be in ratio scales; many seem to be integer values with extremely wide ranges. Also, it would help to know why you want to "select the 3 features with the most significant differences." Although this often seems to be a good way to proceed, it typically isn't. For example, it would miss a feature that is related to the distribution between classes only when the relations of other features to the classes are also taken into account. $\endgroup$
    – EdM
    Sep 27, 2015 at 19:40
  • $\begingroup$ According to page 27 in this book (Data Mining: Concepts, Models, Methods, and Algorithms), "Continuous variables are represented in large data sets with values that are numbers — real or integers". "The difference between these two scales (interval and ratio scales) lies in how the 0 point is defined in the scale". $\endgroup$ Sep 28, 2015 at 13:35
  • $\begingroup$ With regards to why I want to select the most significant features, it is because they all represent a type of features & capture information not directly related to the class I am trying to train a classifier for. I have already trained a classifier (SVM) for it with the standard set of features & by adding the most significant features from the other type of features, I can answer the question of whether the inclusion of features from this other type will improve or degrade the performance of classifiers. Adding all significant features from this other type isn't an option due to some issues $\endgroup$ Sep 28, 2015 at 13:46

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.