International Journal of All Research Education & Scientific Methods

An ISO Certified Peer-Reviewed Journal

ISSN: 2455-6211

Latest News

Visitor Counter
12289121746

Beyond Generic Preprocessing: A Dataset-Adapt...

You Are Here :
> > > >
Beyond Generic Preprocessing: A Dataset-Adapt...

Beyond Generic Preprocessing: A Dataset-Adaptive Framework for Misleading Information Detection in Social Media

Author Name : Krunal D. Vyas, Dr. Nilesh Gambhava

DOI: https://doi.org/10.56025/IJARESM.14112683

 

ABSTRACT Machine-learning detectors of misleading information on social media are highly sensitive to how raw text is prepared, yet preprocessing is rarely documented, tuned or evaluated across different kinds of data. This paper proposes a dataset-specific preprocessing framework that combines field selection, advanced text cleaning, lemmatisation, word n-gram TF-IDF representation, class balancing and grid-search tuning, together with bias (sentiment and political) and credibility (source and speaker) feature engineering. The framework is evaluated with five classifiers—logistic regression, decision tree, random forest, multinomial naive Bayes and passive aggressive—on three public corpora that differ in text length and class balance: Kaggle Fake and Real News articles (44,877 records), FakeNewsNet headlines (23,196) and LIAR statements (12,787). On the article corpus the proposed pipeline raised accuracy from 86.91%, 88.77% and 79.79% for a baseline pipeline (cleaning, stop word removal, TF-IDF) to 99.42%, 95.16% and 99.80% for logistic regression, naive Bayes and random forest respectively. Class balancing improved all fifteen model–corpus pairs, by up to 10.9 percentage points. The best accuracies were 99.80% on articles, 86.79% on headlines and 75.68% on statements, showing that text length and corpus type limit what preprocessing alone can achieve. A review of fifteen recent studies (2021–2025) shows that preprocessing, class balance and evaluation protocols are inconsistently reported, which motivates the framework. A hybrid machine-learning and deep-learning ensemble that builds on the proposed preprocessing is outlined for future work.