DOCOMO's cold-start fix: a Tweedie-powered AI that predicts from day one, with almost no data

NTT DOCOMO announced on Monday in Tokyo the development of a new AI technology, the Dual-view Adaptive Retrieval-augmented Tweedie model, that can make highly accurate predictions even when only limited historical data is available. It targets one of machine learning's oldest headaches: the cold-start problem, where an AI cannot make reliable predictions or recommendations because there is simply not enough history to learn from — the situation every new service, store, or product launches into.

DOCOMO says the model will let service providers offer accurate AI-powered predictions and recommendations from the day a service launches, without waiting for large amounts of new data to accumulate. That could substantially accelerate how fast new businesses come to life.

Why Gaussian training rules fail#

The first of the model's two key features is statistical. Conventional AI training rules use the Gaussian distribution, which assumes data points cluster symmetrically around the mean in a familiar bell curve. That works fine for tidy data — and breaks on the messy kind: data with many zero values, or where peak and off-peak periods behave nothing alike, like social-media engagement or foot traffic past a train station at rush hour versus midnight.

DOCOMO's model instead trains on the Tweedie distribution, a probability distribution flexible enough to represent complex real-world data with lots of zeros and heavy variation. In the company's framing, swapping the Gaussian assumption for Tweedie is what lets the model learn appropriately even when the data is unevenly distributed.

Borrowing from neighbors#

The second feature is retrieval. When historical data for the prediction target is thin, the model looks for nearest neighbors — similar cases that carry useful information — and automatically learns from shared characteristics like location, time, and other attributes. Think other stores in the same area, or other products in similar categories. The model borrows what it can from those neighbors and folds it into its predictions.

First target: digital billboards#

The concrete example DOCOMO gives is digital out-of-home (DOOH) advertising — digital signage and electronic billboards in transportation hubs, outdoor sites, and commercial facilities. The model should be able to predict the number of impressions generated by advertising on newly installed signage from day one, even in locations with wildly fluctuating foot traffic like major urban train stations. That lets ad-slot prices be set and sales begin soon after the screens go up, rather than after months of traffic measurement.

There is a consumer angle too: the company says end users will be able to receive content and information suited to their interests from the moment they start using a service, rather than enduring a cold, generic onboarding period while the system learns who they are.

RecSys acceptance and field trials#

The research has a peer-reviewed stamp: a paper describing the model has been accepted for presentation at the 20th ACM Conference on Recommender Systems (ACM RecSys 2026), a leading international conference on recommendation technology. Acceptance at RecSys signals novelty and demonstrated performance, though real-world commercial performance still has to be proven in the field.

That field test is coming. DOCOMO says it will evaluate the model's effectiveness in field trials with DOOH businesses in Japan and overseas by March 2027, with the goal of supporting a global commercial deployment. DOCOMO — Japan's leading mobile operator with more than 91 million subscribers — frames this as part of its push beyond mobile services into broader technology solutions, and cold-start prediction is the kind of capability that travels: any industry that launches new products into data-poor territory has the same problem.