NTT DOCOMO, INC. announced today the development of a new AI technology, the Dual-view Adaptive Retrieval-augmented Tweedie model, that enables highly accurate predictions even when only limited historical data is available. The model solves the "cold-start problem," which occurs when AI cannot make reliable predictions or recommendations due to insufficient historical data—for example, when a new service is launched.
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 hosted by the Association for Computing Machinery (ACM). This acceptance recognizes the model's novelty and performance.
Quick Intel
When launching a new service, or shortly after expanding an existing service into a different field or region, there is often insufficient historical data for AI-based predictions. This can interfere with AI-powered recommendation features and advertising-performance forecasts—the cold-start problem—resulting in providers having difficulty offering recommendations or tailoring advertisements to user interests, thus making it harder to retain users and develop new businesses.
The model addresses this challenge through two key features.
First, the model uses the "Tweedie distribution," a statistical-probability distribution that can flexibly represent complex real-world data, including data with many zero values or substantial variation. Specifically, this distribution is used in the rules used for training AI models. Conventional AI training rules use the "Gaussian distribution," which assumes that data points are mostly concentrated around the mean, forming a symmetric, bell-shaped curve, a natural pattern in which values close to the mean occur more frequently, while values farther away occur less often. But AI can struggle to learn effectively if conditions are complex, such as when data differs significantly between peak and off-peak periods, as in the case of social media engagement, or when data varies widely. However, the model uses the Tweedie distribution to appropriately train models and make predictions, even when data is unevenly distributed.
Second, the model uses "nearest neighbors"—similar cases that provide useful information when historical data for the target being predicted is limited. It automatically identifies and learns from common characteristics, such as location, time and other attributes, in data from these nearest neighbors, including other stores in the same area and other products in similar categories. The model then incorporates this information in its predictions.
The model will enable service providers to offer accurate AI-powered predictions and recommendations from the day a service launches, without waiting for large amounts of new data to accumulate. This could help to substantially accelerate the launch of new businesses.
For example, in DOOH advertising, the model will be able to predict the number of impressions generated by advertising on newly installed digital signage on the first day—even in locations with substantial fluctuations in foot traffic, such as major urban train stations. These predictions will be made with limited historical data, enabling advertising-slot prices to be set and sales to begin soon after installing the signage.
The model will also enable end users to easily receive content and information suited to their interests from the moment they begin using a service, improving the overall service experience.
Going forward, DOCOMO will evaluate the model's effectiveness in field trials with DOOH businesses in Japan and overseas by March 2027, aiming to support its global commercial deployment.
About NTT DOCOMO
NTT DOCOMO, Japan's leading mobile operator with over 91 million subscribers, is one of the global leaders in 3G, 4G and 5G mobile network technologies.
Under the slogan "Bridging Worlds for Wonder & Happiness," DOCOMO is actively collaborating with global partners to expand its business scope from mobile services to comprehensive solutions, aiming to deliver unsurpassed value and drive innovation in technology and communications, ultimately to support positive change and advancement in global society.
