Forecasting and analyzing technology development trends with self-attention and frequency enhanced LSTM

Abstract

Analyzing and forecasting technology development trends is of significant importance for formulating research and development (R&D) strategies. Existing research focused on analyzing historical trends of technology development or utilizing link prediction to forecast future technology interactions, thereby providing decision support for formulating R&D strategies. However, these methods rely on expert experience or fail to produce predictive trend insights, thus lacking objectivity and effectiveness. To address these issues, we start with predictions of technology interaction intensity trends to understand future technology interactions, thereby providing decision support for R&D. Specifically, we represent technology interactions using the co-occurrence of classification codes and developed a Self-Attention and Frequency Enhanced Long-Short Term Memory (SAFE-LSTM) model to predict the future connection strengths of classification codes, thereby constructing the landscape for future technology interactions. We trained this model on the American patent dataset and compared it with several typical machine-learning methods. The results indicate that the SAFE-LSTM model achieved significant advantages in single- and multi-step predictions. Building on this foundation, we further analyze technology development trends, yielding valuable insights. This study provides researchers with more comprehensive and predictive insights, supporting the integration with additional analytical methods to offer more robust decision-making support for R&D, thereby contributing to the future competitiveness of enterprises.