I want to build feature vectors from data of my test set, which contains profiles of people. I always want to compare two profiles to each other.

Thus my features are:
- Same surname ∈ {undefined, yes, no}
- age delta ∈ {undefined, x | x ∈ Z}
- number of same interests ∈ N
- genders ∈ {(male, female), (male, male), (undefined, male), ...}
- number of common friends (i think this should be normalized by the total number of friends both profiles have) ∈ N

I want to use this feature vectors labeled with 1, -1 to learn classifying a relation between two profiles with a SVM or k-nearest neighbours. I think I should binarize the feature vectors somehow, but I am not shure what is the best way.

My ideas are:
- Just transform the values into binary representation
- Use One-hot encoding
- Split the gender feature into two features: gender_A, gender_B
- Normalize the common friends value by dividing through the absolute of the difference of number of friends for each profile plus one
- Don't normalize the common friends value, just add more features for #friends_A, #friends_B

What do you think would be the best solution or what could I do instead?
Can anyone help me?

  • $\begingroup$ It seems that you have graph values, probably from a social network. I would add some edges or vertex feature like centrality, betweeness, etc... $\endgroup$ – Scratch Mar 26 '14 at 14:51
  • $\begingroup$ That's correct. Centrality and betweeness already are another project. Actually I'm only not sure about the representation of the feature vector. $\endgroup$ – Thorben Mar 26 '14 at 22:08

As for the number of common friend i suggest using the Jacard index. Its basically the ratio between shared friends both friends.


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.