2
$\begingroup$

I am an intern at mobility data company and a Master's candidate in Statistics. I am researching about driving score which is based on a driver's driving habit. We have trip data which contains the distance, quick acceleration(seconds), quick stop, steering angles, and so on. I have read some related articles and papers but some of them contain the skills that I cannot handle such as genetic programming.

I would like to know which ML skills could be used for this unsupervised learning problem(maybe?). Below is an example of our data set.

Driver ID | Trip time | Distance | Harsh acceleration | Quick Stop | ...

1              60 mins     1 mile       180 seconds         7 times  ...

2             30 mins      0.3 mile    10 seconds           2 times ...

My goal is making a driving score based on the dataset for each driver. The scale does not matter. It could be 0-100 scale or classifiers such as poor, bad, normal, good, perfect. The problem I'm struggling with is that I have to create the target value(driving score). I guess that unsupervised learning could be a hint but I am not pretty sure about it. I welcome any kind of advice or source! Thank you so much in advance.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

Since you do not have labelled data, you'll have to use a lot of domain knowledge. Also, predicting a class (poor, bad, normal, good, perfect) might be easier than predicting a score (0-100). The approach you could use is:

  1. Apply clustering on your dataset.
  2. Try to find the optimal number of clusters for your dataset using the Elbow Method. Suppose you get 3 optimal clusters.
  3. Now comes the part for applying domain knowledge. Try to analyse the clusters by picking up individual examples from the cluster and using domain knowledge identify which category (here, good, bad, average) you can classify them into.
  4. After analysing some examples from the clusters, you can assign the label you get to the entire clusters as the examples within a cluster are similar.

The above method makes some assumptions but can be an interesting experiment for your dataset.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thank you for your advice. I only have tried 2 variables clustering. Do you mean that dimension of N>2 is possible? Or dimensionality reduction such as PCA would be required before clustering? $\endgroup$
    – Hun Cho
    Jun 13, 2019 at 9:45
  • $\begingroup$ Yes, of course, it is possible. You can try PCA if you feel that some features are redundant. $\endgroup$
    – bkshi
    Jun 13, 2019 at 10:00

Your Answer

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

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