NTT DOCOMO Unveils AI Model That Predicts With Limited Data
NTT DOCOMO has developed a new AI technology that can make predictions even when only limited historical data is available. Called the Dual-view Adaptive Retrieval-augmented Tweedie model, the technology is designed to address the AI “cold-start problem,” where new services, products or locations lack enough past data for conventional models to make reliable predictions.
DOCOMO said the technology could support recommendation systems, advertising-performance forecasts and other applications where businesses need AI predictions soon after launching a new service. A research paper describing the model has also been accepted for presentation at ACM RecSys 2026.
Model Tackles the AI Cold-Start Problem
AI systems generally become more effective as they collect historical data. This creates a challenge when a company launches a new service or expands an existing product into a new region, where there may be little information available about user behaviour or demand.
DOCOMO's model is designed to make use of limited information rather than waiting for a large historical dataset to build up. The company says this could allow AI-powered recommendations and forecasting to operate from the early stages of a service launch.
Tweedie Distribution Handles Uneven Data
One part of the model uses the Tweedie distribution, a statistical approach capable of handling data with large variations and many zero values. DOCOMO says this is useful for situations where activity can differ sharply between peak and off-peak periods.
Traditional training approaches can struggle when data does not follow a relatively stable distribution. By incorporating the Tweedie distribution into the model's training process, DOCOMO aims to improve predictions from irregular and limited datasets.
AI Learns From Similar Cases
The second component uses nearest-neighbour information to supplement limited data. The model identifies similar cases based on characteristics such as location, time and other attributes, then incorporates information from those cases into its predictions.
For example, a newly opened store with little historical information could draw on data from similar stores in the same area. Likewise, a new product could use information from products in related categories to improve its initial predictions.
Digital Advertising Is a Key Use Case
DOCOMO sees potential applications in digital out-of-home advertising (DOOH). The model could estimate advertising impressions for newly installed digital signage from the first day of operation, including in locations such as busy railway stations where pedestrian traffic can fluctuate considerably.
Those predictions could help advertising operators determine slot prices and begin selling advertising space without waiting months for sufficient historical data.
Field Trials Planned Through March 2027
DOCOMO plans to evaluate the technology through field trials with DOOH businesses in Japan and overseas by March 2027. The company ultimately aims to support commercial deployment of the technology globally.
The acceptance of the underlying research at ACM RecSys 2026 gives the technology an academic testing and dissemination pathway, while the planned field trials will determine how effectively it performs in practical business environments.
