New Media Public Opinion Heat Perception and Trend Prediction based on Machine Learning Algorithm

As internet technology advances, new media platforms have become pivotal for information dissemination and public sentiment expression, critically influencing governance, corporate reputation, and social stability. Traditional public opinion analysis methods often struggle with temporal dependency modeling and multi-feature integration, limiting predictive accuracy. Motivated by the need for dynamic, high-precision sentiment forecasting, this paper proposes an LSTM-based time series model incorporating self-attention mechanisms to weigh temporal influences. By integrating keyword frequency and emotional polarity within sliding time windows, the framework captures contextual nuances while enabling nonlinear mapping via a dense layer. Experimental validation demonstrates stable performance (0.74-0.95 across metrics), outperforming conventional approaches. Contributions include a novel hybrid architecture and adaptive optimization, advancing public opinion analytics. The paper is structured as follows: methodology, experimental design, results, and implications for practice.