Synerise on the podium of the AI Challenge by

5 min read

The Synerise AI team, led by Jacek Dąbrowski, took 2nd place in the WebTour 2021 Challenge organized by

The rivalry of the leaders was close, with Synerise losing marginally to specialists from Nvidia - a company valued at around $350B. The competition focused on a multi-destination trip planning problem, which is a popular travel industry scenario. The goal of this challenge was to use a dataset based on millions of real, anonymized accommodation reservations and come up with a strategy for making the best recommendation for their next destination in real-time. 

Trip planning usually involves the need to synchronize various products—means of transport, accommodation, tourist attractions, etc. In addition to considering factors such as time and availability, the decision-making process is not based solely on rational and objective criteria. As a result, it is difficult to provide relevant recommendations to visitors on a travel site promptly.

Contestants had to propose specific methods and techniques for finding information and recommendations, taking into account mobility, arising preferences, emotions, as well as the travel and tourism environment in the COVID-19 and post-covid era.

The solution prepared by the Synerise team consisting of Jacek Dąbrowski (Chief AI Officer), Konrad Gołuchowski (AI Lead Scientist), Barbara Rychalska (AI/NLP Researcher) and Michał Daniluk (Research Scientist) turned out to be one of the best of all those submitted. The model was based on Cleora and EMDE—two algorithms entirely developed by our team that are used daily in Synerise's production systems. 

Cleora is a graph embedding algorithm which is significantly faster than competitors (~200x faster than DeepWalk and 4x-8x faster than PyTorch-BigGraph) and it is an input method for EMDE (creates embedding vectors which are fed to EMDE). EMDE is a compression method for sets of vectors spanning very high dimensional spaces. It creates small and efficient representations of multimodal item sets (e.g. user sessions in recommendation) and allows for predicting relevant items from these compressed structures. EMDE is essentially a recommender system. Here we applied it to the space of city-vectors, to represent a user’s past travel history. In the contest we have proven that EMDE is great at representing sequential data (sequences of cities a user is visiting).

"At Synerise AI we have decided to abandon the well-trodden path of small, incremental improvements and explore new and unexpected directions of research. We accept the high odds of failure which come with exploration, knowing that risk defines the value of the ultimate goal. Our approach has led us to the formulation of a novel research direction unifying differential geometry, statistics, and deep learning. Based on this, we have created a suite of unique tools which comprise of a fast, accurate and resource-efficient recommender system whose accuracy is amongst the world's best. At the same time, we care greatly about the practical value of our inventions, aiming for them to be used by actual people and businesses on real-world data to solve real-world problems. The Booking Challenge was especially interesting as the nature of the problem was in many ways different than regular recommendation tasks on which our suite scores top results. We are thankful for the distinction and the possibility to confirm the usefulness of our approach to this completely new domain of sequential route planning.”

Barbara Rychalska, AI/NLP Researcher at Synerise

Synerise is currently one of the most famous and innovative Polish companies operating in high technology sectors crucial to the economy, such as big data and artificial intelligence. The company boldly invests in its own data processing products and artificial intelligence solutions to completely change the modern approach to data management. 

In July 2020, the Synerise AI team won the prestigious competition organized by Rakuten Institute of Technology and presented two papers during the International ACM SIGIR Conference on Research and Development in Information Retrieval held in Xi'an, China.