Tokyo – October 12, 2026 -- NTT DOCOMO, INC. has developed an AI model that delivers accurate predictions from the first day of a new service, and it will test the technology in field trials with digital out-of-home (DOOH) advertising businesses in Japan and overseas by March 2027.
The Dual-view Adaptive Retrieval-augmented Tweedie model targets the "cold-start problem." This occurs when AI cannot make reliable predictions or recommendations because historical data is insufficient, for example after a service launch or an expansion into a new field or region. A paper on the model has been accepted for presentation at the 20th ACM Conference on Recommender Systems (ACM RecSys 2026).
DOCOMO replaces Gaussian training rules with the Tweedie distribution
Conventional AI training rules rely on the Gaussian distribution, which assumes data clusters symmetrically around the mean. According to DOCOMO, AI struggles to learn under that assumption when data differs sharply between peak and off-peak periods, as with social media engagement, or when it varies widely.
The new model applies the Tweedie distribution in its training rules. The distribution can represent complex real-world data, including data sets with many zero values or substantial variation, so the model can train and predict even when data is unevenly distributed.
Nearest-neighbor retrieval fills gaps where target data is thin
The model also draws on "nearest neighbors," meaning similar cases that supply useful information when history for the prediction target is limited. It automatically identifies common characteristics such as location and time. Examples of neighbors include other stores in the same area and other products in similar categories. It then feeds that information into its predictions.
DOOH advertisers could price new signage on day one
In DOOH advertising, the model is designed to predict first-day impressions for newly installed digital signage, including at locations with large swings in foot traffic such as major urban train stations. Those forecasts would let operators set advertising-slot prices and begin sales soon after installation, rather than waiting for data to accumulate.
DOCOMO says the approach could substantially accelerate new business launches. It also expects end users to receive content and information matched to their interests from the moment they start using a service.
DOCOMO plans to evaluate the model in the DOOH field trials by March 2027 and aims to support global commercial deployment.