Discovery Challenges

MoReBikeS: Model Reuse with Bike rental Station data
Brief Description

Adaptive reuse of learnt knowledge is of critical importance in the majority of knowledge-intensive application areas, particularly when the context in which the learnt model operates can be expected to vary from training to deployment. This challenge therefore focuses on model reuse and context change.

The challenge is carried out in the framework of historical bicycle rental data obtained from Valencia, Spain. Bicycles are continuously taken from and returned to rental stations across the city. The data consists of time series describing hourly availability of bikes at each station; information on weather and (local) holidays is also provided. The challenge motivation is based on the fact that, while we may have had the opportunity to learn and tune good models for old stations with historical data, we do not always have the same amount of data for new stations. With that in mind, participants will receive, in addition to limited data for the new stations, a large number of trained models for old stations. The task will be to make predictions (3 hours ahead) with regard to the number of bikes available for these new stations and within the next months. This situation fluctuates considerably depending on the time of year, the station's location, etc. The key point here is that by using models from other stations that have been learnt from data spanning more than one year, better predictions can be made for the new stations. In the end, this challenge aims at promoting the reusability of models rather than retraining a different model again and again each time the context changes.

 

Challenge website

 

On Learning from Taxi GPS Traces
Brief Description

Electronic taxi dispatch systems are in wide use today. These systems have replaced the traditional VHF-radio dispatch by installing mobile data terminals in the taxis, which typically provide GPS localization information and taximeter state. In the last couple of years, the broadcast-based radio messages for service dispatching were replaced by unicast-based messages between the taxi central and the selected vehicle.

In most cases, taxi drivers operating through an electronic dispatch system do not indicate the final destination of the ride. In some cases, particularly when the demand for taxis is higher than the taxi availability, the closest taxi to a particular location is exactly the taxi that will end its current ride at that location. While in broadcast-based radio dispatching this was not a problem, in unicast-based electronic dispatching it becomes a problem, given that most drivers do not indicate the final destination of their current ride. To improve the efficiency of electronic taxi dispatching systems it becomes important to be able to predict the final destination of busy taxis. The spatial trajectory of a busy taxi could provide some hints on where it is going. Similarly, given the taxi id it might be possible to guess its final destination based on the regularity of pre-hired services. In a significant number of taxi rides (approximately 25%), the taxi has been called through the taxi call-center, and thus the passenger’s telephone id can be used to narrow the destination prediction based on the historical ride data of such telephone id.

The main goals within this challenge are the following:

a)     To predict the final destination of each taxi ride (namely, its coordinates – latitude, longitude);

b)    To estimate the travel time of each taxi ride since the beginning of the ride.

Challenge website