What Can Online Forums Tell Us? A Customized Deep Learning Approach to Examining Online Posts from E-Cigarette Consumers

The emergence of advanced deep learning algorithms, such as ChatGPT and other large language models (LLMs), has accelerated rapidly in recent years. Building on this progress, the present study introduces a novel deep learning algorithm—Transformer-Enhanced E-cigarette Satisfaction Analyzer (TEESA)—to analyze consumer-generated social media posts about e-cigarettes from online forums. TEESA is grounded in an eight-dimensional framework, the PUT-E-SPEC model, which was developed through a systematic literature review on e-cigarette consumer satisfaction. This deep learning method integrates the PUT-E-SPEC framework to measure various dimensions of users’ experiences with e-cigarettes. To validate the PUT-E-SPEC model, we conducted a randomized experiment involving 900 participants recruited online via Amazon Mechanical Turk. Participants were randomly assigned to nine groups: one control group and eight treatment groups, each treatment corresponding to one of the PUT-E-SPEC dimensions. The preliminary result confirms the effectiveness of PUT-E-SPEC. Built on the RoBERTa model (Liu et al., 2019), TEESA employs attention-based contextual embeddings and multi-step reasoning to capture nuanced aspects of product satisfaction. Compared to benchmarks such as ChatGPT and other LLMs, TEESA demonstrates significant improvements in interpretability and contextual understanding, achieving a joint accuracy of 0.47—translating to over 0.90 accuracy across all eight dimensions—ensuring high precision in assessing consumer sentiment. Our consumer-generated training data was sourced from reputable platforms such as Reviews.io and ConsumerAffairs.com, covering the period from 2014 to 2024. Additionally, we collected over 300,000 textual posts from the “electronic cigarette” subreddit on Reddit.com—one of the most active online forums for e-cigarette discussions—for sentiment analysis using the PUT-E-SPEC framework. By applying TEESA to forum data on two e-cigarette brands (Brand A and Brand B), we found that TEESA-generated sentiment scores were significantly correlated with the brands’ differing market performances across the eight PUT-E-SPEC dimensions. This study makes two key contributions: (1) technically, we developed a novel customized deep learning algorithm tailored for analyzing user-generated content; and (2) empirically, our findings may assist public health experts and policymakers in understanding consumer perceptions and guiding decisions using deep learning analyses of large-scale online forum data for public health issues, such as e-cigarettee.