Abstract
Strong in-domain performance in fake-news detection does not necessarily imply robustness across datasets, sources, or claim distributions. This study investigates a comparatively compact CNN–BiLSTM architecture for text-based fake-news detection using repeated random-seed evaluation, controlled classical and neural baselines, computational profiling, post hoc explanation analysis, and cross-dataset evaluation. The architecture combines a trainable embedding layer, one-dimensional convolution, and two stacked bidirectional LSTM layers, with 1,532,273 trainable parameters. Across five prespecified random seeds, the model achieved mean Accuracy and Macro-F1 of (99.79%) and (99.79%), respectively, on ISOT, and (97.84%) and (97.84%) on the fixed tweet-level TruthSeeker protocol. The TruthSeeker results require caution because approximately (99.94%) of test tweets are associated with underlying statements already represented in the training partition. On binary LIAR2, the model achieved mean Accuracy of (76.47%) and Macro-F1 of (70.32%), demonstrating substantial dataset dependence. Controlled comparisons showed that no single architecture dominated across all datasets and metrics. Strict zero-shot transfer from ISOT to the PolitiFact and GossipCop subsets of FakeNewsNet resulted in substantial performance degradation. A matched title-to-title transfer experiment improved PolitiFact performance but did not resolve the weak transfer to GossipCop, indicating that input-field mismatch accounts for only part of the external-domain degradation. LIME showed variable local-surrogate fidelity, while LIME and SHAP provided complementary local attribution patterns. The results support the CNN–BiLSTM as a compact and competitive text-based alternative while highlighting the importance of controlled baselines, overlap analysis, cautious interpretation of post hoc explanations, and strict external evaluation.