Kaggle Days Dubai Competition: Optimizing police patrols


It is a tradition, to have the second day solely for the offline competition on Kaggle Days. For such a purpose we ask businesses and government divisions for the cooperation. Everyone benefits this way:

  • For the Sponsors – Competition on Kaggle Days is a great opportunity to show the ways in which they are engaged in AI field; directly interact with the best Data Scientists around the whole world and even recruit talents.
  • For the Participants – it is an opportunity to learn new topics, challenge their skills, team up with colleague data scientists and learn from others, and, also, interact with the people ‘behind’ the problem.

For this edition, we were happy to cooperate with Dubai Police. It is one of the most progressive police forces in the world. To maintain this prestigious title they constantly challenge themselves and explore new areas for improvement. That’s why they invest in AI solutions that are helping all Dubai citizens.

Participants coding

To enable fair and solvable competition we selected the topic for the competition in Dubai to be: “Predicting the number of incidents within different time slots in more than 300 regions in Dubai”.

We loved this topic especially because this was an obvious but necessary step to build a system that would allow the Police to react in a shorter time for every incident. Yes, we are talking about being on the spot within less than 10 minutes – wow!

Apart from the competition – Police has asked us for help with two other vital topics for brainstorm, which you can find child abuse prevention. road fatalities reduction, and competition.

For the competition, Dubai Police has provided a large amount of data for the last 4 years containing incidents, which after anonymization was available for the participants. The dataset was simple at first glance but has a huge potential for feature engineering.

Competition topic

The goal of the competition was to predict the number of incidents in specific areas of Dubai. Participants were provided with historical data of incidents by areas for the last 4 years, and additionally with weather, temperature and public holidays datasets. Algorithms were tested on December 2018 using RMSE metric.

You can check out the competition and the leaderboard here.

From left to right: Walter Reade, Maria Parysz and Pawel Jankiewicz present algorithm that clusters regions of Dubai

NEW mentoring model

To make the competition process more productive, for the first time we have asked Kaggle Grandmasters to be mentors of four randomly selected teams and help them in their fight for the first place.

Kaggle Grandmaster Abhishek Thakur with the team card

During previous editions, Mentors were able to support all Teams. Dubai Kaggle Days was the first time we introduced this new mentoring model and everybody loved it – it turned out to be very successful and well received by the participants, as well as mentors themselves. With this approach, Mentors were also competing, but this time as Mentors.

This image has an empty alt attribute; its file name is CimW26NtDEZ-uZZaB3FpxiDdKuaqz480n9_LU1iM7FjQzyHrFS9QiXW1ZJf4EcmryZL8iZ-aqHpsdo7ZWnFGyUnOhoaCksYhzLKzDwCMRq0p0Cm2nP4g9v-pUU5cAhXpS2u6tadM
Kaggle Grandmaster Giba mentors his team
This image has an empty alt attribute; its file name is ccZbgoHSHt93Gn3-unpeDNpLkZDSf0X0zxln2Rhtdl6_d3ycARPdfNnJmeV1UfsEBV1HwE-6kADaVgjpTUS3m0RKRuWDKrBhS4C9b7fnCKz6543Cq7ZbfiItvMSzL-5eux2atMcC
Kaggle Grandmaster Abhishek Thakur consults one of his teams

Twenty-six teams competed within 9,5 h in our offline competition. Kaggle Days team provided participants with weather and landmarks data additionally to the main datasets.

After 9 hours everybody was nervous and curious to see the leaderboard change after changing Publice Leaderboard to Private one. To everyone’s surprise, there was the biggest shake-up in our history: team from the 5th place popped up to the 1st, and 2nd place solution raised even more: 12 places up.

The winners are:

  1. ‘better than ever’: Vit D, Paul Kochetkoff and Richard
  2. ‘Crime_Oracles’:  Ajay Pratap Singh Pundhir, Martin Petrov and Georgy Regas
  3. ‘STERMEDIA’: Piotr Giedziun

We also announced the best Mentors, – Pavel Ostyakov and Gilberto Titericz. Two winning teams (the 1st and 2nd places) were mentored by Pavel Ostyakov, while the third place took the team mentored by Giba.

The winners

We have asked our winners about some insights to the winning solutions. Mainly LightGBM and Linear Regression were used with the extensive feature engineering. You can read about second’s place solution here.

On top of the winning solutions LogicAI Team, lead by Pawel Jankiewicz have prepared something extra,  – an algorithm that clusters regions in Dubai in such a way, that can lead to a shorter response time of the police forces.

After the intense work on the competition, everyone was curious about the winner’s solutions. At the end of the day, all together we went to the bar to celebrate the winners and discuss their approach and what they did to win the competition.

————–

Kaggle Days Dubai took place in Dubai World Trade Centre on April 30 – May 1, 2019. Please, find detailed article here.

Kaggle Days are a series of offline, non-profit events for Data Science enthusiasts and Kagglers, created by LogicAI in partnership with Kaggle. Kaggle Days events aim to provide an opportunity to learn from Kaggle Grandmasters and network within the community. We gather worldwide Data Science community to encourage learning and sharing experience.

LogicAI is a data science consulting company providing end to end solutions for business problems using advanced analytics and machine learning models. We are authors of Dubai competitions and we provide services in those fields like: predicting number of incidents, optimizing routes for police patrols, detecting child abuse in social media and reducing number of fatalitites.