1
$\begingroup$

I have a dataset with 3 columns - user_id (mostly unique), timestamp of first opening of the app by the user(unique and inside a range of few hours - I believe this is the time of launch of the app) and the device model with brand name (like HTC Desire X). I have been cracking my head to find any clustering or correlation or anything at all but this seems very useless imo. The given features seem to be too less information. Any ideas as to what I might do to derive useful inference from this data? (I apologize if this violates any community guidelines)

$\endgroup$

1 Answer 1

1
$\begingroup$

It may not seem a lot of information but there are ways of extracting potentially interesting results. You could, for example, not only consider the individual timestamps but also their distribution.

It all depends the details of your dataset but you could try to, for example, first create some sort of "adoption curve". This distribution then represents the amount of users using a certain brand opened the app from first to last day.

In this way you could try to cluster on the distribution of timestamps instead of timestamps which on their own might not be that informative. Clustering these distributions could be done on their exact form, but I would rather go for mean, variance, min, max, ... etc.

I can imagine that there might be some relation between certain brands and certain kinds of users that might be more likely to try out new apps, resulting in characteristics in their distribution.

Some inspiration

The following example from a course on Big Data & Analytics I took used taxi-arrival data (arrival timestamp and coordinates) from the San Francisco Bay Area to infer neighbourhoods. The arrival times or coordinates on their own might not tell us much.

The coordinates are therefore first clustered into 750 geographic regions. For every region, a "taxi profile" is made that represents the distribution of arrival times in the timespan of a week. These profiles are then clustered into 6 clusters which have similar "taxi-profile", i.e. where the pattern in taxi drop-off times is similar. After averaging out over all taxi-profiles inside each cluster, this results in the following 6 "taxi-profiles" that represent a certain type of area:

enter image description here

Using these profiles, different region like for example, the financial district (region 3) and residential areas (region 2) can be identified and represented on a map:

enter image description here

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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