DaCFake: Hybrid Fake News Detection

• The paper introduces a divide and conquer approach that merges handcrafted linguistic cues and word vectors, achieving outstanding accuracy (96-98%) across multiple datasets.
• DaCFake is a hybrid fake news detection model that decomposes text into linguistic and context-based features, effectively capturing stylistic and semantic nuances.
• The model employs rigorous feature selection and ensemble classifiers, ensuring robustness and generalizability for detecting fake news in diverse social media environments.

DaCFake is a fake news detection model for social media that applies a divide and conquer strategy to combine content and context based features extracted from news articles. The framework separates feature engineering into handcrafted linguistic features and word vector features, integrates the resulting representations after preprocessing and feature selection, and then evaluates a range of machine learning classifiers. In the reported experiments, DaCFake was evaluated on three datasets—Kaggle, McIntire + PolitiFact, and Reuters—and achieved accuracy rates of 97.88%, 96.05%, and 97.32%, respectively, with ten-fold cross validation used to enhance robustness and accuracy (Jain et al., 22 Aug 2025).

1. Problem framing and overall design

DaCFake is motivated by the claim that the proliferation of fake news on social media requires rapid and automated detection methods because traditional fact checking methods are often too slow to prevent the dissemination of false information. The framework is explicitly described as a hybrid ML system that combines advanced feature engineering with robust machine learning algorithms for early and accurate fake news detection across diverse datasets (Jain et al., 22 Aug 2025).

Its central organizational principle is a two-phase divide and conquer strategy. In the Divide Phase, the input text is decomposed into two main feature categories: Linguistic Features (LF) and Word Vector Features (WVF). In the Conquer Phase, the selected LF and WVF representations are concatenated into a unified input for downstream classifiers. The workflow also includes preprocessing, feature selection using Pearson correlation, an 80:20 train/test split, and evaluation of multiple ML models, with ensemble methods reported as the strongest performers.

This design places DaCFake in a line of fake news detectors that rely on explicit feature construction rather than exclusively end-to-end neural sequence models. A plausible implication is that the method is intended to preserve interpretability at the feature level while still benefiting from distributed lexical representations.

2. Feature decomposition in the divide phase

The linguistic branch extracts a comprehensive set of 80 features, organized into nine groups (Jain et al., 22 Aug 2025):

• LF1: Character-level: number of characters, digits, letters, uppercase, whitespaces, special character frequency.
• LF2: Word-level: number of words, short words $<4$ chars, average sentence and word lengths, word frequency bands, type-token ratio.
• LF3: Syntactic: frequency of punctuation and function words.
• LF4: Structural: number of lines, sentences, paragraphs, sentences per paragraph, average words per paragraph, quoted content, URLs.
• LF5: Content: frequency of content words.
• LF6: Readability: established readability indices including Flesch, Kincaid, SMOG, and Coleman-Liau.
• LF7: Uncertainty: certainty words, tentative words, modal verbs.
• LF8: Specificity: adjectives, adverbs, affective terms.
• LF9: Verbal Non-Immediacy: self-reference, group-reference, and first-, second-, and third-person pronouns.

Feature extraction uses LIWC (Linguistic Inquiry and Word Count 2007), the textstat Python library, and custom Python scripts. The reported feature inventory indicates that DaCFake treats stylistic, structural, and discourse-level regularities as discriminative signals rather than relying only on lexical identity.

The context-based branch uses word embeddings derived from CBOW or Skip-gram (SG), described as Word2Vec variants. The text states a vocabulary size of 1,000, and the workflow diagram labels the embeddings as “Word Embeddings (CBOW/Skip-gram, 1000 dim)”. For this branch, the preprocessing pipeline removes stopwords, URLs, punctuation, emojis, dates, and numbers, and then applies stemming and lemmatization (Jain et al., 22 Aug 2025).

The article text therefore distinguishes two complementary representational regimes: LF captures handcrafted textual regularities, whereas WVF captures contextual lexical information through dense vectors. This suggests that DaCFake is designed to mitigate the weaknesses of either representation used in isolation.

3. Feature integration, selection, and classifier stack

After LF and WVF extraction, DaCFake concatenates the two vectors to form the final feature representation. The combined vector is then normalized, with the paper stating that normalization scales features to a common range, typically $[0,1]$. Redundancy reduction is performed through Pearson correlation-based feature selection: features with correlation $|r| > 0.7$ are considered redundant and removed (Jain et al., 22 Aug 2025). The correlation is given as

$$\rho(X, Y) = \frac{\mathrm{cov}(X, Y)}{\sigma_X \sigma_Y}.$$

The training pipeline is summarized in the paper’s Algorithm 1 as follows: preprocessing removes redundant or missing data and performs label encoding; feature engineering extracts LF and WVF; feature union combines LF and WVF and standardizes them; train-test splitting uses an 80:20 division; and several ML classifiers are then evaluated with 10-fold cross validation.

The classifier set includes ensemble models—Gradient Boosting (GB), AdaBoost, Random Forest (RF), and ExtraTrees (ET)—alongside Support Vector Machine (SVM), Logistic Regression (LR), Naive Bayes (NB), and K-Nearest Neighbors (KNN). The paper reports that ensemble methods showed consistent superior performance (Jain et al., 22 Aug 2025).

A common misconception about systems of this kind is that they incorporate social context in the sense of user metadata or propagation graphs. DaCFake does not do so. Its “content and context based features” are textual: handcrafted linguistic cues plus word-vector representations. The paper explicitly lists as a limitation that no user metadata, propagation patterns, or network structure are employed.

4. Datasets, preprocessing, and evaluation protocol

DaCFake was evaluated on three article-level datasets whose compositions are explicitly reported (Jain et al., 22 Aug 2025).

Dataset	Composition	Best reported accuracy
Kaggle	20,800 articles; 10,387 TRUE / 10,413 FAKE	97.88%
Reuters	44,898 articles; 21,416 TRUE / 23,482 FAKE	97.32%
McIntire + PolitiFact	6,755 articles; 3,382 TRUE / 3,373 FAKE	96.05%

All datasets contain article title, text, and label (“fake” or “true”). Preprocessing removes duplicates, missing values, irrelevant content, and standardizes labels. The paper describes phased cleaning for LF and WVF, reflecting the different requirements of handcrafted-feature extraction and word embedding construction.

Validation uses 10-fold cross-validation, stated to be stratified for reliability. The reported metrics are Accuracy, Precision, Recall, and F1-score, with the formulas given as

$\text{Precision} = \frac{TP}{TP+FP},$

$\text{Recall} = \frac{TP}{TP+FN},$

$$\text{F1} = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision}+\text{Recall}},$$

$$\text{Accuracy} = \frac{TP + TN}{TP + FP + FN + TN}.$$

The implementation environment is reported as Python 3.6, with SciPy, scikit-learn, NLTK, and Textstat. The hardware specification is Intel i9, 64GB RAM, RTX 3080 (Jain et al., 22 Aug 2025).

5. Empirical performance and reported textual indicators

Across all three datasets, the strongest configuration was LF + Skip-gram (WVF) with Gradient Boosting, which yielded the headline accuracies of 97.88% on Kaggle, 97.32% on Reuters, and 96.05% on McIntire + PolitiFact (Jain et al., 22 Aug 2025). The paper further states that using only LF or only word vectors reduced accuracy into an approximate range of 91–96%, whereas integrating both was optimal. Precision, recall, and F1-score are described as similarly high.

The comparative discussion in the paper reports that DaCFake outperforms many recent fake news detection systems, including WELFake (SVM, 96.73%) and some BERT-based models (~97.45% on Kaggle), while noting that these are typically evaluated on fewer datasets or show lower generalization. Within the scope of the reported experiments, the strongest evidence for DaCFake’s effectiveness is therefore cross-dataset consistency rather than performance on a single benchmark.

The paper also identifies several LF indicators as particularly informative. It reports that fake news tends to use more exclamation points and question marks, more third-person pronouns, more URLs, shorter sentences, higher type-token ratio, and less quoted content (Jain et al., 22 Aug 2025). These observations are consistent with a stylistic-diagnostic view of fake news detection in which rhetorical intensity, referential style, and document structure are predictive.

6. Significance, limitations, and prospective extensions

The paper attributes several strengths to DaCFake: generalizability across heterogeneous datasets, rich feature engineering, strong use of ensemble models, and early detection capability. In methodological terms, the system’s distinguishing feature is not a novel neural encoder but the systematic combination of handcrafted LF and WVF under a structured divide and conquer pipeline (Jain et al., 22 Aug 2025).

At the same time, the paper lists clear limitations. Computational complexity remains nontrivial because feature extraction over large datasets, especially with Word2Vec models and ensemble methods, can be resource-intensive. Feature dependence is another limitation: the most effective features may vary by language and news domain. The framework also excludes user metadata, propagation patterns, and network structure, making it a text-centric detector rather than a full social-media misinformation analysis system. Finally, model quality remains dependent on dataset quality, including possible noise and labeling errors.

The future directions proposed in the paper are correspondingly specific: incorporation of user and temporal features, improved scalability and efficiency for very large streaming datasets, and cross-lingual or multimodal extensions beyond English text alone (Jain et al., 22 Aug 2025). A plausible implication is that DaCFake serves as a modular baseline for hybrid misinformation detection systems in which explicit linguistic signals remain central even when additional modalities are later introduced.