Artificial intelligence is often described as a technology that improves as more data arrives. That strength becomes an awkward weakness at the moment a business may need prediction most: the first day of a new store, product, service or advertising location, when there is almost no history to learn from.
NTT DOCOMO says it has developed an AI method aimed directly at that problem. Announced on September 28, the Dual-view Adaptive Retrieval-augmented Tweedie model—DART—is designed to produce useful predictions when historical data for the target is scarce. The approach combines a Tweedie probability model, suited to uneven data with many zero values, with retrieval of similar “nearest-neighbor” cases. A paper on the method has been accepted at the 20th ACM Conference on Recommender Systems, RecSys 2026.[1]
DOCOMO’s first proposed proving ground is digital out-of-home advertising, or DOOH. Imagine a new digital screen installed in a major railway station. On opening day, there is no history showing how many impressions that exact screen will generate. DART is intended to borrow information from similar locations and contexts—other sites in the area, comparable time periods and related attributes—to estimate performance before its own data history exists.[1]
DOCOMO plans field trials with DOOH businesses in Japan and overseas during the fiscal year ending March 2027, with an eye toward broader global deployment.
The paradox of AI at launch
Recommender systems usually learn from behavior: what people viewed, clicked, bought, watched or ignored. Mature services may have millions of interactions. A new user or new item has none. The resulting “cold-start problem” has been one of recommender systems’ persistent challenges for decades.
DOCOMO has encountered it before. In a 2023 technical-journal article on content recommendation, its researchers described two forms of cold start: new users whose behavior has not accumulated and new content with no click history. That system used content metadata such as titles and genres so that even never-clicked items could receive recommendation scores.[2]
DART pushes the problem beyond content recommendation toward a more general prediction setting. If the target itself has little history, the model searches for relevant structure elsewhere.
First idea: stop assuming the data forms a neat bell curve
The first major component is the Tweedie distribution. DOCOMO contrasts it with Gaussian assumptions commonly used in modeling, where observations cluster symmetrically around an average. Many commercial data sets do not behave that way.[1]
Pedestrian traffic may be quiet for hours and then surge during commuting peaks or an event. Social-media engagement can contain huge numbers of low or zero responses and a tiny number of viral outliers. Advertising impressions can have similar skew.
Tweedie distributions are a family of exponential-dispersion models in which variance follows a power relationship with the mean. Depending on the power parameter, the family connects to familiar distributions including Gaussian, Poisson, gamma and inverse Gaussian. In the important range between 1 and 2, a Tweedie model has a point mass at zero while also supporting positive continuous values—a useful structure for data in which “nothing happened” is common but nonzero outcomes vary widely.[3]
The statistical idea predates modern AI. Tweedie models have long appeared in areas such as insurance and rainfall modeling because those fields also contain many zeros and highly variable positive outcomes. DOCOMO’s contribution is to incorporate that statistical structure into an AI model for cold-start prediction.
Second idea: when your own past is missing, borrow from similar cases
DART’s other major component is retrieval of nearest-neighbor examples. A new store may have no sales history, but nearby stores do. A new product may lack demand history, but similar products in the same category have it. DOCOMO says the model identifies common features such as place, time and other attributes in those neighboring examples and incorporates them into its prediction.[1]
The logic resembles an experienced operator opening a new branch. The manager cannot consult “last year at this store,” because the store did not exist. Instead, the manager looks at surrounding foot traffic, comparable stores, day of week, time of day and local demographics. DART formalizes that kind of analogy inside the model.
DOCOMO’s public release does not provide a full mathematical account of what the “dual-view” label represents, so Japan.co.jp does not infer one beyond the disclosed architecture. The published description centers on the combination of Tweedie-based learning and adaptive retrieval of similar cases.
Why digital outdoor advertising is a logical first application
DOOH is not a random demonstration choice. DOCOMO is already building businesses around audience estimation and location data.
In April 2026, the company announced that Vie BOARD CORPORATION, its joint venture with Vietnam’s DATVIETVAC group, would launch a digital-signage network in Ho Chi Minh City. DOCOMO described the project as its first entry into overseas DOOH and said it would provide its DOCOMO Data Science technology, which uses overseas GPS data to estimate audience size and demographic attributes such as age, gender and interests.[4]
DART can therefore be understood as an additional layer on top of a business DOCOMO is already operating: not merely measuring an established advertising location, but estimating a new one before much local history has accumulated.
If a new screen’s impressions can be estimated from day one, an operator may be able to price inventory and begin selling it sooner. That turns prediction accuracy into something more concrete: shorter time between installing an asset and monetizing it.
“Highly accurate” still needs a numerical context
There is an important limitation in the public information. DOCOMO describes DART as enabling “highly accurate” prediction and says its paper was accepted at ACM RecSys 2026. But the September 28 release does not disclose the benchmark data sets, competing baselines, error metrics or numerical improvement rates used to support that characterization.[1]
Japan.co.jp therefore does not attach an independent percentage improvement to the technology. Nor should conference acceptance be confused with commercial validation. Peer review is meaningful evidence that the research was judged worthy of presentation; it does not establish profitability, reliability across industries or performance under every real-world data shift.
DOCOMO’s own plan reflects that distinction: field trials in Japan and overseas come next.
The conference itself is marking 20 years
ACM RecSys began in 2007 and holds its 20th conference in 2026. This year it returns to Minnesota, where the first meeting was held. The main sessions run from September 29 through October 1 in Minneapolis, with workshops and tutorials around them.[5]
The field has broadened dramatically over those two decades. Early recommender-system research was heavily associated with collaborative filtering—using patterns among users and items to say, in effect, “people who liked this also liked that.” The 2026 program spans explainability, fairness, multimodal recommendation, conversational systems, generative and agentic methods, evaluation beyond accuracy and industrial deployment.
Cold start, however, remains. If anything, its business importance grows as more organizations attempt to introduce AI into products that do not yet have the historical data of a large established platform.
DOCOMO has been moving from communications data toward business data
NTT DOCOMO began operations in July 1992. It launched i-mode in 1999, turning the mobile handset into a mass-market gateway for internet services, introduced commercial 3G FOMA service in 2001 and one of the early LTE services in 2010.[6]
Its present business is substantially wider than mobile connectivity. DOCOMO’s corporate description includes consumer communications, smart-life businesses such as content and marketing solutions, and enterprise communications. As of March 31, 2026, DOCOMO reported 9,876 employees at the company and 53,780 across its group.[7]
That evolution helps explain why an operator is publishing recommendation and forecasting research. For a telecom company focused only on network carriage, AI might primarily optimize radio resources. For a group selling advertising, content, payments and enterprise solutions, prediction itself becomes part of the product.
Small-data AI may matter as much as ever-larger models
The current AI narrative often focuses on enormous models trained on enormous data sets. Enterprise reality frequently looks different. A new factory product has few defect examples. A newly opened branch has no local demand history. A rare equipment failure, by definition, offers few training cases. A business entering a new country cannot simply manufacture years of historical behavior.
This is why few-shot learning, zero-shot learning, transfer learning, synthetic data and retrieval-based methods are attracting sustained attention. An NTT technical review published in 2026, discussing industrial image recognition, identified few-shot and zero-shot methods as important ways to reduce the burden of collecting hundreds or thousands of labeled examples for every new defect type.[8]
DART belongs to that wider shift, but with a particular focus: numerical outcomes that are sparse, skewed and difficult to model, combined with transfer of information from similar historical cases.
The next important evidence will come from deployment
Three questions stand out for the coming DOOH trials.
First, how accurate is DART when a location truly has almost no history—not merely a smaller training sample drawn from a mature data set? Second, how well does neighbor retrieval travel across cities and countries, where pedestrian behavior and advertising markets may differ? Third, as a new target accumulates its own data, does the model smoothly shift from borrowed information toward direct evidence?
There are also governance questions. Advertising and recommendation systems depend on data about people and contexts. Accuracy alone is not enough; privacy, transparency, bias and the definition of what counts as a “similar” neighbor can all influence outcomes.
If DART performs in operational trials as DOCOMO expects, its most valuable contribution may be time. A business would not need to wait months for enough history before using AI. It could begin with informed analogies, then learn from its own evidence as that evidence arrives.
That is a more modest claim than an AI that knows the future. For new businesses, it may also be a much more useful one.
