NTT DOCOMO Builds AI System for Precise Forecasts with Minimal Data
NTT DOCOMO, INC. has built a new artificial intelligence system designed to make precise forecasts using minimal historical information. The Dual-view Adaptive Retrieval-augmented Tweedie model targets the persistent cold-start problem that routinely undermines digital advertising campaigns and new business launches.
Breaking Past the Limits of Gaussian Distributions
Launching an unfamiliar service or expanding into new regional markets triggers a predictable operational roadblock. Insufficient historical data starves predictive algorithms, crippling automated recommendation engines and paralyzing advertising-performance forecasts.
Traditional AI training rules rely on the Gaussian distribution. This statistical approach assumes data points cluster symmetrically around a mean, forming a bell-shaped curve. Complex real-world environments routinely shatter this assumption. Social media engagement fluctuates wildly between peak and off-peak periods, and enterprise sales data often exhibits extreme variance or heavy zero-value clusters. Standard models struggle to learn effectively under these jagged conditions.
Using Tweedie Distribution and Nearest Neighbors
To bypass these limitations, DOCOMO’s new model incorporates the Tweedie distribution. This probability distribution flexibly represents complex real-world data featuring numerous zero values or substantial variations. By applying the Tweedie distribution to AI training rules, the technology maintains high predictive accuracy even when incoming data is unevenly distributed across operational timelines.
The architecture also deploys nearest neighbors analysis to bridge information gaps. When target data is scarce, the system automatically identifies and learns from common characteristics—such as geographical location, time stamps, and categorical attributes—pulled from similar products or neighboring stores. This contextual synthesis allows service providers to deploy functional, data-backed recommendations from day one.
Academic Validation at ACM RecSys 2026
The academic validation of the underlying research is already secured. A paper detailing the model has been accepted for presentation at the 20th ACM Conference on Recommender Systems, known as ACM RecSys 2026, hosted by the Association for Computing Machinery. This acceptance acknowledges the novelty and performance metrics of DOCOMO’s architecture.
Field Trials and Urban Transit Advertising
Practical enterprise applications extend immediately into digital out-of-home advertising. Newly installed digital signage in high-traffic urban train stations typically lacks historical foot-traffic data. With the new model, operators can accurately predict impression counts on the very first day of installation. Advertisers can price inventory and close sales without waiting months for performance metrics to accumulate.
End users benefit concurrently through immediate personalization. Consumers receive tailored content matching their expressed and inferred interests from their initial interaction with a platform, driving higher early engagement rates. DOCOMO plans to evaluate the model’s commercial viability through rigorous field trials with digital out-of-home businesses in Japan and international markets by March 2027.