NTT DOCOMO has developed a new AI model designed to make accurate predictions when historical data is scarce, addressing one of the persistent challenges in machine learning: the cold-start problem. The Dual-view Adaptive Retrieval-augmented Tweedie model combines a probability distribution suited to uneven real-world data with information retrieved from similar cases, potentially allowing new services to deploy AI recommendations and forecasting from launch.
AI systems often become more useful as they accumulate data. That creates a problem for businesses that want to use machine learning from day one.
NTT DOCOMO is targeting that AI cold-start problem with a new model that can make predictions when historical information for a new service, product, location or market is limited. The telecommunications company says its Dual-view Adaptive Retrieval-augmented Tweedie model could eventually support applications ranging from recommendations to digital out-of-home (DOOH) advertising.
The research paper describing the model has been accepted for presentation at ACM RecSys 2026, the 20th ACM Conference on Recommender Systems. That gives the work a peer-reviewed research venue, although commercial performance will still need to be established through field testing.
Cold-start problems occur when an AI system has insufficient historical observations to identify reliable patterns. A new digital service, newly opened retail location or recently introduced product may have little or no behavioral history. Conventional recommendation and forecasting systems can therefore struggle precisely when businesses need them to work.
DOCOMO’s approach combines two techniques to address the limitation.
The first is the Tweedie distribution, a statistical distribution capable of modeling data with substantial variation and large numbers of zero values. DOCOMO says this makes it better suited to certain real-world prediction problems than conventional training approaches based on the Gaussian distribution.
That distinction matters because many business datasets do not resemble a neat bell curve.
Advertising impressions, for example, can vary dramatically between peak and off-peak periods. Foot traffic at a railway station can change depending on the time, day, location or surrounding events. A model that assumes relatively consistent variation around an average can struggle when observations are highly uneven.
The Tweedie-based approach is designed to account for that kind of distribution during training.
The second component is retrieval from nearest neighbors. Rather than relying exclusively on historical observations for the exact target being predicted, the model identifies similar cases and incorporates their characteristics.
For a new retail location, that could mean learning from comparable stores in the same area. For a newly introduced product, similar products or categories could provide additional signals. Location, timing and other attributes can become part of the information used to identify those comparable cases.
This is where the model’s “dual-view” approach becomes particularly relevant to machine learning infrastructure and AI development frameworks. The system effectively combines statistical modeling of the target data with information retrieved from related examples, allowing the model to compensate for the lack of direct historical observations.
The potential commercial impact is significant if the approach performs consistently outside research environments.
Businesses frequently delay AI-powered services because they need enough behavioral data before predictive models become reliable. A system capable of producing useful forecasts immediately could shorten that ramp-up period and make AI more practical for newly launched services.
DOCOMO’s proposed DOOH application illustrates the idea.
Digital signage operators often need to estimate how many people will see an advertisement before they have accumulated sufficient historical information about a newly installed screen. A digital sign located in a major railway station could experience large variations in pedestrian traffic depending on time and conditions.
DOCOMO says its model could predict advertising impressions from the first day using limited historical data. That could allow operators to establish advertising-slot prices and begin selling inventory sooner rather than waiting for months of observations.
The same principle could apply to recommendation systems.
A new service typically knows little about a customer who has just signed up. Instead of treating that user as an isolated data point, an AI system could use information from similar users, products, locations or behaviors to generate an initial recommendation.
That makes the work relevant to Generative AI Technologies, Enterprise AI Applications and AI Automation Platforms, even though the underlying technique is more closely associated with predictive machine learning than generative AI.
It also reflects a broader shift toward retrieval-enhanced AI. Modern AI systems increasingly combine learned models with external information rather than expecting the model itself to contain every relevant signal. In DOCOMO’s case, retrieval is being used not to augment a large language model, but to provide additional evidence for statistical prediction.
That distinction is important.
Large language models such as those developed by Google, Microsoft and Amazon are driving much of the current AI market, but many enterprise problems still require specialized predictive models. Demand forecasting, advertising measurement, pricing and recommendation systems often depend on structured data, probability distributions and domain-specific signals.
The challenge is proving that these specialized models can deliver reliable results under genuinely cold-start conditions.
DOCOMO plans to conduct field trials with DOOH businesses in Japan and overseas by March 2027, with the longer-term objective of supporting global commercial deployment.
If those trials validate the research results, the technology could provide a practical way for businesses to reduce the data advantage enjoyed by established products and services.
The broader significance is straightforward: AI does not always need years of historical data to become useful. By combining flexible statistical modeling with information from similar cases, DOCOMO is attempting to make prediction systems more effective at the moment businesses arguably need them most—when a new product, service or market has just begun generating data.
Market Landscape
The AI industry has largely focused on scaling models with more parameters, data and computing resources. DOCOMO’s research addresses a different problem: how to make machine learning useful when data is scarce.
Cold-start prediction is particularly relevant to recommendation engines, retail analytics, advertising, pricing and new-market launches. Retrieval techniques can supplement sparse target data, while specialized statistical models can account for distributions that are poorly represented by conventional assumptions.
The approach also complements rather than competes directly with LLM-centric AI. Enterprises increasingly use different model architectures for different workloads, combining generative AI with predictive machine learning, retrieval systems and domain-specific analytics.
DOCOMO’s planned field trials will be important in determining whether its research can translate into reliable commercial forecasting across different locations, industries and data conditions.
Top Insights
- NTT DOCOMO’s new model targets the AI cold-start problem by making predictions when a new service has little historical data.
- Tweedie-based training is designed to handle uneven distributions, zero-heavy datasets and substantial variation in real-world observations.
- Nearest-neighbor retrieval allows the model to learn from comparable locations, products and other cases when target data is limited.
- DOOH advertising could benefit by forecasting impressions for newly installed digital signage before substantial historical traffic data exists.
- Field trials planned through March 2027 will test whether the research can translate into commercial AI prediction services.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
