A Comparative Performance Analysis of Transformerbased Consumer Sentiment Models and Traditional Machine Learning Algorithms for Predicting Sustainable Purchase Intent in Crossborder Digital Markets

Cross-border e-commerce is exploding and increasingly intersecting with a global consumer trend toward sustainable and eco-friendly products. The success of these heterogeneous, multilingual digital markets relies on the ability to accurately perform sentiment analysis on user-generated content to predict consumers’ purchase intent. Conventional machine learning algorithms that rely on rigid lexical features and term frequencies are often inadequate to capture the nuance of the complex, contextual discourse of cross-cultural sustainability. In this paper, we present a thorough comparative analysis of the performance of legacy machine learning algorithms (Support Vector Machines, Random Forests) and state-of-the-art Transformer-based architectures (mBERT, XLM-RoBERTa) for the prediction of sustainable purchase intent. When tested on a large multilingual set of e-commerce reviews and social media interactions around the world, the Transformer models show a remarkable advantage in contextual understanding. The XLM-RoBERTa architecture achieved the highest F1-Score of 94.2%, which is substantially higher than the conventional baselines, especially in non-English texts that contain localised idioms and complex eco-jargon. Transformer models are more computationally intensive, but they provide cross-border digital retailers with highly actionable and accurate predictive intelligence to achieve sustainable market expansion.