I have a dataset of what apps users downloaded and I try to estimate these users' gender based on what apps they downloaded, using machine learning algorithm. However, what kind of features of the apps should I focus on? As far as I know, app category plays a big role. Do Google Play and Apple App Store estimate users' gender based on what apps they download?
If you have data on how many downloads have been done for each category, you can use that data to perform logistic regression, where your target variable would be 0/1 classification of male/female. This will form your benchmark result and then you can improve upon the results. Especially look for feature selection, because there might be categories of apps which have nothing to do with target. You should also try bi variate profiling of variables with target to see if there is good correlation with target. Try using some advanced techniques once you have done some data exploration. KNN, Random Forest and Neural network for classification might give good results. Also cart might work fantastically here.
Regarding the features, except for the app category, I would try:
- app size (in MBs) (i.e. avg of size of the downloaded apps, sum of different sizes, etc.) This would be important, since games tend to be huge, whereas productivity apps tend to have a smaller footprint.
- number of app downloads (again, avg and/or sum of the download counts for all the apps the user has downloaded).
- count of downloaded apps
- number of different categories (that gives an estimation of the variance of the categories the subject user inspects)