NTT DOCOMO unveils AI cold start model for sparse data

NTT DOCOMO has developed the Dual View Adaptive Retrieval-Augmented Tweedie model to tackle the AI cold start problem when little historical data is available. It combines a Tweedie distribution with nearest-neighbor learning and targets recommendation systems and digital advertising. Field trials with digital out-of-home advertising firms in Japan and overseas are planned by March 2027.

• Model uses a Tweedie distribution instead of Gaussian for uneven data
• Nearest-neighbor learning draws on similar locations and products
• Field trials with DOOH advertisers in Japan and abroad by March 2027
• Paper accepted for presentation at ACM RecSys 2026