Cinimex Makes The Podium At Aeroclub Challenge 2023
Cinimex team took the second place at a hackathon organized by Aeroclub company. Made up of two stages, the hackathon gathered 140 teams. 12 teams that qualified for the final stage, based on their metrics, got a chance to pitch their project to a jury of experts. Cinimex won the silver prize among all solutions submitted to the challenge.
Aeroclub Challenge 2023 theme was B2B services in air travel, and namely, offering B2B travelers plane tickets based on customer’s needs and limitations.
Historically, buying airplane tickets for business trips means that contracting a travel agent that will select top-5 from about 200 different options based on customer’s filters and needs. The challenge set by the hackathon was to automate this process, but not only identify the best travel options, but also simulate agent’s behavior.
Challenge participants had to keep in mind that the tools needs to range flight combinations with different characteristics from the most relevant to the least convenient, and that this model has to be scalable and admit new information and agent activity.
As a starting point, participants were given the historical data on the top-5 options selected from all the ones that pass the filters. Historical data included information about the route, travel conditions and traveler’s level.
Cinimex’s team did an in-depth study and added the background information with airport data and a good number of other parameters such as flight time, possible jetlag due to different time zones, and many others. The new data brought about a new challenge for the model to build – how do we split the historical data for training, testing and validating the results of modeling in the best possible way?
Cinimex’s team chose the Stratified Group Split approach, which made sure that the teaching and validation samples corresponded to the distribution in the values of historical routes. Besides, the team studied the most relevant approaches and models for the solution, and chose to go with CatBoost library from Yandex. The team also used Optuna framework to optimize the set of hyper-parameters.
The most important parameters turned out to be: flight duration, number of layovers, and ticket price, while luggage allowance and air carrier ended up as numbers 5 and 12 in the ranking. The project culminated in a model that ranks travel options appropriately, and about 2 out of the top-5 options produced by the model are viewed as relevant by agents.
The key value of the proposed solution is modularity of the process for adding external data and possibility to integrate the solution with existing systems. The solution can be helpful in a number of sectors: plane ticket, hotel, accommodation and transport rental aggregators; air travel and hotels; call centers and marketplaces.