NTT DOCOMO Develops AI Model for Predictions When Historical Data Is Limited

NTT DOCOMO announced a new AI technology called the Dual-view Adaptive Retrieval-augmented Tweedie model. NTT DOCOMO has announced an AI model designed for prediction tasks where historical data is limited. The technology, called the Dual-view Adaptive Retrieval-augmented Tweedie model, targets a problem that appears whenever a new service or situation does not have enough past observations for conventional forecasting methods.

Prediction without years of history

The method uses a Tweedie distribution to handle data with many zero values and uneven variation.

The cold-start problem is common in technology products. A new service may have many users but little historical behavior, while a new category can lack the long record that machine-learning systems normally rely on. Waiting years for enough data is not practical when a company needs predictions during the early stage of a service.

NTT DOCOMO’s research release provides the technical context for the prediction method including retrieval and a Tweedie-based model.

DOCOMO said a paper describing the model was accepted for presentation at ACM RecSys 2026.

• NTT DOCOMO announced a new AI technology called the Dual-view Adaptive Retrieval-augmented Tweedie model.
• NTT DOCOMO has announced an AI model designed for prediction tasks where historical data is limited.

NTT DOCOMO is addressing a common problem in machine learning: new services and products often have too little historical data for conventional recommendation or prediction models to work well.

A cold-start model can be useful when the system has information about related items or users but lacks direct historical observations. The research challenge is to use that limited information without producing predictions that appear more certain than the evidence supports.

DOCOMO’s approach combines retrieval with a Tweedie-based prediction model. The important point is not simply the model name, but the attempt to make useful predictions from sparse evidence by bringing additional information into the forecasting process.

The technique could be relevant to telecommunications and other businesses where services, plans and user behavior change quickly. Its practical value will ultimately depend on how accurately it performs on new datasets and whether the approach transfers beyond the conditions used in the research.