Experimental Dataset for Fake News Detection with CLIP-Guided Cross-Modal Hierarchical Attention

It contains four files:(1) Comparative experiment results (CGCHA.csv): covering the Weibo and PHEME datasets. The Weibo dataset includes results comparing CGCHA with 11 baseline models; the PHEME dataset includes results comparing CGCHA with 9 baseline models, totaling 18 unique baselines across both datasets.

Evaluation metrics include Accuracy, Precision, Recall, and F1 for both Fake and Real categories. Eight baselines (EANN and MFAN on Weibo; KDMRD, MKEMN, MIN, CMAC, MFAN, and CAF-ODNN on PHEME) report Accuracy only, with other metrics unavailable in the original literature.(2) Ablation study results (CGCHAablation.csv): also covering both Weibo and PHEME datasets, each containing 12 experimental groups corresponding to the full CGCHA model and 11 ablation variants (-CLIP, -BV, -ATT, -CV, -CB, -GF, -Layer1, -Layer2, -Layer3, Swap-Layer, Para-Fusion), designed to validate the contributions of key components including the CLIP encoder, co-attention mechanism, gating fusion, and the three-layer progressive structure.(3) Per-epoch training logs (weibo_accuracy.txt, twitter_pheme_accuracy.txt): recording the full training process of the CGCHA model on the Weibo (2,160 epochs) and PHEME (1,037 epochs) datasets, including training/test loss, accuracy, precision, recall, F1, and per-class metrics for each epoch.