How to classify samples with different features?

Question

Assuming we are considering following classification problem:

We have a dataset containing the time when a user call a taxi in one day, but different users call the taxi different times. For example:

User  Time 
1   7:00    11:00   13:00   16:00           
2   7:01    12:30   17:30               
3   7:30    9:30    11:00   11:45   12:50   16:00   18:00
4   6:30    23:00                   
5   9:00    11:00   13:00   17:00           
6   9:00    17:30                   

Let say, User 1,2,5,6 can be classified as one group and User 3 and User 4 are the other two group. And in the User 1,2,5,6 is looks like User 1,2 can be viewed as a subgroup and User 5,6 may be another subgroup.

If I want to this kind of classification, what is the possible method? Should I apply some preprocess procedure to transform the data into a square shape? Or there are methods can trade this kind of data without preprocessing?

I am not sure this problem is well-posed. Any discussion is welcomed.

Answer 1

For this kind of data I would recommend to convert the data into integer values to make them easier to compare, but this step could also be omitted. E.g. you could have a function like:

time = 60 * hour + minutes

Next question would be how to compare the data. Is it more important for you to categorize the users by the time he/she calls a taxi or how often per day he/she calls a taxi.

One approach to preprocess the data could be to split a 24 hour day into time periods of 30 minutes or 1h and rearrange the data like:

time  ...  7:01-8:00  8:01-9:00 ...
User1 ...  0          0         ...
User2 ...  1          0         ...
User3 ...  1          0         ...
User4 ...  0          0         ...
User5 ...  0          1         ...
User6 ...  0          1         ...

Where the values are the taxies called in this time; could also be just a boolean value. By setting the box sizes different properties can be observed like in the morning, afternoon, night or something different. Also the times must not start at 0 am, the could also start at the minimum time which occurs and go to the maximum, which would remove rows that are completely zero.

With this new matrix we can apply a distance function like the euclidean distance.

Based on the resulting distance matrix on could apply some clustering technique like Principal Component Analysis, Hierarchical Clustering, k-Means or some other clustering method you prefer.

Important for the results is the size of the time-step you choose. A bigger time-step would lead to more similarities between the users. To round this up, I would say, that this kind of preprocessing leads to an identification of users, that call taxies to equal times.