SENADA: A Stacked Ensemble Learning for Native Advertisement Detection in Electronic News

Abstract

Native advertisements are increasingly common as hybrid commercial content embedded within digital news platforms. Although these ads can enhance user engagement, they often elicit negative reactions due to their concealed commercial intent. A key challenge in detecting native advertisements lies in their lack of explicit labeling, which can mislead readers and compromise the perceived credibility of news content. Prior studies have struggled to accurately identify native ads because of their inability to capture the implicit characteristics embedded in editorial-like text, such as positive sentiment, persuasive language, one-sided perspectives, and references to specific products or companies. To address this issue, this study proposes a deep learning architecture called SENADA (Stacked Ensemble for Native Advertising Analysis). The model integrates BERT-based contextual representations, ensemble learning strategies, and a BiLSTM with an attention mechanism. This design enables the model to effectively identify subtle patterns and contextual cues associated with native advertisements. In addition, we introduce a newly constructed Indonesian-language dataset of news articles, annotated based on the four implicit characteristics of native ads. The dataset was collected from six major e-news portals and manually labeled by trained annotators. Experimental evaluations, based on confusion matrix analysis, demonstrate that SENADA achieves an accuracy of 0.93, along with high values for precision, recall, and F1-score. The model is compared against several well-established baselines, including CNN, BiLSTM, BiGRU, and transformer-based architectures, which were selected based on their widespread use and proven effectiveness in prior text classification studies. The proposed model consistently outperforms these baselines, particularly in minimizing misclassification of borderline content, defined as news articles that exhibit some but not all native ad characteristics, such as persuasive tone without explicit product mention or positive sentiment without clear promotional framing. These borderline cases were validated through manual annotation and cross-checking among annotators to ensure reliability. These results confirm the model’s effectiveness in detecting native advertisements that exhibit complex and implicit textual patterns. While the model shows promising results, future work is needed to reduce architectural complexity and training time. Further research could explore explainability methods and cross-lingual adaptation to extend the model’s applicability across languages and media contexts.