JOAI 2026: Winning Solution Leverages Transformer and BiLSTM for 3rd Place
research#transformer📝 Blog|Analyzed: Feb 20, 2026 15:30•
Published: Feb 20, 2026 15:18
•1 min read
•Qiita AIAnalysis
Element138 achieved a remarkable 3rd place in the JOAI 2026 competition, showcasing the power of combining time-series feature engineering with advanced neural network architectures. The solution's elegance lies in its simplicity, utilizing just two models based on Transformer and BiLSTM. The use of data augmentation and pseudo-labeling further enhanced the model's performance.
Key Takeaways
- •The solution employed a combination of Transformer and BiLSTM models.
- •Time-series feature engineering, including 1st and 2nd order differences, played a crucial role.
- •Data augmentation, pseudo-labeling, and pre-training techniques contributed to the improved score.
Reference / Citation
View Original"By using time-oriented feature creation (1st and 2nd order differences, etc.) and metadata embedding, and using Transformer and BiLSTM as backbones, good results were obtained."