A novel three-step approach to forecast firm-specific technology convergence opportunity via multi-dimensional feature fusion

A team of researchers has published a novel AI framework designed to solve a critical problem in innovation strategy: predicting which technologies will converge to create new opportunities for specific companies. While technology convergence (TC) is a major driver of innovation, existing methods primarily operate at the industry level, offering little actionable insight for individual firms. The new approach, detailed in an arXiv preprint, tackles this by fusing three previously underutilized dimensions of patent data—bibliometric, network structure, and textual features—using attention mechanisms to create rich, multi-dimensional representations at the International Patent Classification (IPC) level.

The core of the method is a three-step pipeline. First, it extracts and fuses the multi-dimensional patent features. Second, a two-stage ensemble learning model, equipped with strategies to handle imbalanced data, identifies potential TC opportunities at the IPC level. Finally, and most innovatively, the system refines these opportunities into actionable, firm-specific recommendations using a large language model (LLM) in a retrieval-augmented generation (RAG) setup, which acts as a judge to evaluate the feasibility and relevance of the predicted convergences. The researchers demonstrated the system's effectiveness by applying it to Zhejiang Sanhua Intelligent Controls, a leading Chinese auto parts manufacturer, successfully forecasting TC opportunities in the high-stakes domains of thermal management for energy storage and robotics.