NTT DOCOMO Develops AI for Better Predictions With Less Data

NTT DOCOMO has announced a new AI technology designed to improve prediction accuracy even when only a small amount of historical data is available. The development addresses one of the long-standing challenges in artificial intelligence known as the cold-start problem, where an AI system has insufficient information to make reliable recommendations or predictions when a new service, product or business begins operation.

The technology is called the Dual-view Adaptive Retrieval-augmented Tweedie model, or DART. According to DOCOMO, the system combines a different statistical approach with information gathered from similar examples. This allows the model to learn from related data even when the target service itself has very little historical information.

The cold-start problem can be particularly important for companies launching new digital services. Traditional AI systems generally become more useful as they collect larger amounts of historical user and business data. A new service, however, may not have enough information during its early days. This can make it difficult to personalize recommendations, estimate demand or predict advertising performance.

DOCOMO says its new approach tackles this challenge in two main ways. First, it incorporates the Tweedie distribution into the AI training process. This statistical method can handle data with large variations, including datasets where many observations are zero but occasional values are very high. Such patterns can occur in areas such as social-media engagement, advertising impressions and changing customer activity.

The second feature involves identifying and using similar cases, sometimes referred to as nearest-neighbor information. When the system has limited information about a particular product, location or service, it can examine related examples and identify common characteristics such as location, time and other attributes. This information can then support predictions for the newer or less-established target.

One potential application is digital out-of-home advertising, or DOOH. Digital billboards and screens located in busy areas can experience major variations in the number of people passing by throughout the day. DOCOMO says its technology could help estimate advertising impressions for newly installed digital signage even before substantial historical data has accumulated.

This could have practical implications for businesses. Advertisers and digital-signage operators may be able to estimate campaign exposure and establish advertising prices earlier instead of waiting for months of data collection. For customers, the technology could also help new services provide more relevant recommendations from an earlier stage.

DOCOMO said the research paper describing the technology has been accepted for the ACM RecSys 2026 conference, an international conference focused on recommender systems. The company plans to conduct demonstration trials involving DOOH businesses in Japan and overseas during fiscal 2026.

The announcement reflects a wider technology trend in 2026: AI development is increasingly moving beyond large language models and content generation toward systems designed for prediction, recommendation and business decision support. Improving AI performance when data is limited could become particularly useful for newly launched products, services and markets.

For companies adopting artificial intelligence, the development also highlights an important issue: having sophisticated AI is not always enough. The quality, quantity and relevance of available data can strongly influence how useful an AI system becomes. Technologies that can work effectively with limited information could therefore become an important part of future AI applications.