Uber Nairobi Ambulance Perambulation Challenge
$6,000 USD
Can you use ML to create an optimised ambulance deployment strategy in Nairobi?
1030 data scientists enrolled, 331 on the leaderboard
ConstructionTransportationHealthPredictionStructuredLocation
Kenya
17 September 2020—24 January 2021
130 days
#2 Winning approach
published 11 Feb 2021, 08:42
edited 2 days later

Dear all,

First of all, we do like to thank Zindi for this wonderfull, challenging and very inspiring challenge and congrats to all winners and competitors.

Here is the summary of my team's winning solution.

Our first intuition was to find best 8 static locations to place along the day, thus 6 ambulance locations to place for each intervall hour 3H of a day, and then optimize it as much as possible the distance according to the scoring function for the leaderboard.

The real challenges for us was:

  • Find a set of crash locations with which to optimize the initials locations.
  • Find the best initial ambulance locations.

Find the best set of crash location for optimization:

Grouping crashes per 3H didn't help cause that leads to overfitting. Thus optimizing according to all the crash locations after removing outliers was better. But that didn't give the real best score.

After some analysis of the weather data, We have taken only those accidents that are in similar weather condition as the time interval in the submission to priorize the test set, and in order to conserve some information from the whole dataset, used kmeans to add some plausible/representative crash locations when k is the best size of the representative crash locations.

Find the best initial locations:

For each interval 3H, we used a combinaison of initial location, all got from kmeans.

All initial locations got by each intervall gave a good score but not the best. After some analysys, we constate a groups of hour that have similar number of crashes, then for each group we set the same initial location. Apart from the initials obtained from each interval, we also used ambulance location obtained from the crash locations for the optimization.

Finally, we minimize the distance sum from the scoring function using gradient descent.

If some of you have some question, we will be happy to answer you in the comment section or in private discussion.

Thanks all,

Alefa @nalysoa, @Joely, @Kajy

Wow. Great work

Thanks,

Your solution is the best since you got the first place.

Thank you very much for sharing!