1
$\begingroup$

I am working with a dataset with about 400 features, all binary (1 or 0). What approach would you recommend? Data set is about 500k records.

$\endgroup$
6
  • $\begingroup$ What's wrong with treating the 0s as 0s and the 1s as 1s? $\endgroup$
    – user253751
    Aug 24, 2021 at 14:51
  • $\begingroup$ Nothing wrong with it, just didn't get me anywhere, so was wondering if maybe there are some rules that say "Binary 1/0 features need to be treated differently than 0-1 range features" $\endgroup$
    – Darko
    Aug 24, 2021 at 14:52
  • $\begingroup$ I think it has more to do with the model you're working with, than with how you input the features. $\endgroup$
    – user253751
    Aug 24, 2021 at 14:58
  • 3
    $\begingroup$ What do you want to do with the data in the first place? $\endgroup$ Aug 24, 2021 at 15:23
  • 2
    $\begingroup$ @Darko Please, edit directly your post to include the details of your problem (because comments are temporary). It's not even clear if you're trying to solve a classification problem or any other problem. In this type of question, it's also important to show us what you have found or tried so far. So, I would suggest that in your edit you also include these details. $\endgroup$
    – nbro
    Aug 24, 2021 at 23:11

1 Answer 1

1
$\begingroup$

Most standard algorithms will work well on binary data, like:

  • Decision trees (and random forest)
  • Nearest Neighbors
  • Neural Networks
  • etc

But your choice depends on many other things, like:

  • What is the expected output?
    • Are you doing classification?
    • Regression?
    • Is it deterministic (same features should always give the same outcome) or stochastic (random factor).
  • What is the nature of the database and relationship between the features?
    1. A 20x20 black-white image.
    2. A phrase embedded as 20 sequences of 20-size-token.
    3. A 400 questions true/false exam.
    • They can all have the same shape, but are very different in nature and would perform better with different algorithms.
    • How disperse / smooth is your data?
    • Do all 400 features have the same importance?
    • How independent are they?
  • How complex the problem really is?
  • How much performance do you really need?
  • How much work and tuning are you willing to put on this?
$\endgroup$
5
  • $\begingroup$ Classification - end result is true/false, deterministic A 400 questions true/false exam is the best description. Unclear on importance of the features. False positives are not an issue, as long as true negatives are minimal to non-existent. How much work and tuning are you willing to put on this? 1-2 weeks full time $\endgroup$
    – Darko
    Aug 24, 2021 at 18:08
  • $\begingroup$ If you could just describe the problem, it would make it easier to recommend something. $\endgroup$ Aug 24, 2021 at 18:17
  • $\begingroup$ If it's like an exam, with a previously expected answer for each feature, a simple neuron network with a threshold would perform fine. If each feature impacts on another, the network must be deeper. If features are highly correlated, you might want to reduce them... There are too many IFs to provide a straight answer. $\endgroup$ Aug 24, 2021 at 18:21
  • $\begingroup$ The features are unrelated. The closest to the scenario is: a user selects if they like or don't like 400 products, and based on that i am trying to determine if they will like the label product. $\endgroup$
    – Darko
    Aug 24, 2021 at 19:02
  • $\begingroup$ If the data is about preference (like or don't like), I would consider a neighbor approach (like KNN). If you have no idea and low previous knowledge, I'd also consider some AutoML approach for setting a quick baseline. $\endgroup$ Aug 24, 2021 at 19:26

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .