RecSys Challenge 2025 is an international competition held as part of the ACM Conference on Recommender Systems (RecSys), a leading international conference in this field. In its most recent edition, a volunteer team of data scientists from Recruit Co., Ltd. and Indeed Recruit Technologies (“rec2”) took first place among 416 teams worldwide. We spoke with Yuki Sawada of Recruit’s Product Development Data Promotion Office about how the team came together, what enabled this result, and what the team learned from their experience.

Backed by a Program That Supports Individual Initiative, a Cross-company Volunteer Team Took on the Challenge

Yuki, who works on using data science to advance matching and product development in the Data Promotion Office, explains: “Within the office, we believe the source of value lies in the capabilities of each data scientist. We have a Cloud Support Program that helps employees take on challenges like this competition and hone their skills.”

Leveraging this program, Yuki called for volunteers, reaching out by saying “If you’re interested, let’s do this together.” Six members from two Recruit Group companies responded and came together to enter the competition.

Sharpening Skills Through External Competitions

According to Yuki, the RecSys Challenge features real-world service data from platforms such as Spotify and X, and asks teams to compete on recommendation quality. Essentially, the competing teams propose solutions to solve the same kinds of problems faced by global tech companies, and their results are evaluated for accuracy.

The theme tackled by team “rec2” was predicting future customer behavior from past on-site activities such as page views, purchases, and cart actions. Their goal was to build user representations, Universal Behavioral Profiles (UBP), that could be applied across multiple prediction tasks to support business decisions. For six tasks, including churn and purchase prediction, the organizers ran a standardized training and evaluation pipeline on a scoring server using each team’s submitted per-user embedding, then compared overall performance.

To win, team “rec2” combined contrastive learning and multi-task learning based on Transformer models, adding rule-based aggregation of behavioral data and stacking the outputs to maximize performance.

For a deeper dive into the technical details, read the Recruit Tech Blog (Japanese only).

Members of the “rec2” volunteer team gathered in the office to watch the end of the competition together

From an International Competition Back to Business Impact

From September 22 to 26, the team went to Prague to join the RecSys 2025 workshop to present their winning approach and to discuss it with other participants.

“Learnings from these kinds of competitions can translate into business outcomes, and an award like this showcases Recruit’s data science expertise on a global stage,” notes Yuki. “Data science is an essential capability for using data to deliver consistently high-quality experiences. Proving our capabilities in a competition that directly tests skills tied to real service improvement is one indicator that Recruit Group can raise recommendation quality to a world-class level.” He adds, “We will continue to sharpen our skills and contribute to advancing matching in our core businesses.”

Team “rec2” after presenting at the Prague workshop (Shugo Takei, Yūhi Nagatsuma, Rintaro Hasegawa, Yuki Sawada, Hiromu Auchi, Kazuhito Yonekawa)

Yuki Sawada

Data Science Department, Data Solutions Unit, HR Product Data, Indeed Recruit Technologies Co., Ltd. Data Science & Machine Learning Engineering Department, Data Technology Unit, Data Promotion Office, Product Development, Recruit Co., Ltd.

After completing graduate school in 2021, Yuki joined DeNA before moving to Recruit in 2025. He has experience in generative image modeling, search, and recommender systems. He currently focuses on data science for HR services at Indeed Recruit Technologies while also contributing to cross-domain knowledge sharing in data science.